Return on nature credits: agriculture

0.1Purpose

This document sets out a method for working out the return a company can expect from buying nature credits in its agricultural supply chain. It explains the question the method answers, the data it uses, how a change in crop yield becomes a change in what a company is worth, and which of its inputs are evidence and which are judgement.

The returns from protecting and restoring nature are documented for whole sectors and countries. There is little public analysis of the return to a single company, and none that measures what nature credits add to it. The method is meant for a company or investor to apply to their own business or client.

The document can be read without knowing the tool. Parts 1 to 6 describe the method. Part 7 says how far it can be trusted and what it leaves out. Part 8 describes what the programme does for nature. The appendices include what is needed to build the model again.

0.2The question

A company buys a crop from a particular landscape. The climate there is changing, and the farms it buys from will grow less. That affects the company's finances and valuation.

The method answers two questions in turn. First, what financial value does an improvement in nature on those farms offer the company and its business model? Second, does buying nature credits give a better return than funding the same work directly?

The answers are given as company value and cash flows, the measures a company uses for capital decisions.

0.3What the method does not do

• It does not forecast prices. Price paths are inputs, chosen by the user and checked against past prices (§3.3, §6.2). The method makes no claim about what a crop will cost in any year.
• It does not put a price on individual nature outcomes, such as tonnes of carbon or numbers of species. The economic return comes from changes in land management that support better nature outcomes. The models of carbon, species and soil each measure something different, so there is no consistent way to put a separate value on each outcome. This method is also not a crediting methodology. Each credit project has its own crediting methodology, and that project is where the actual nature outcomes are measured. Part 8 describes the change in nature.
• It is not an offset scheme. The company pays for a better supply of the crop it buys. The company does not offset any emissions, and no credits are cancelled.
• It does not build its own crop model. How well a crop grows under each climate and each way of farming comes from a published crop model (§2.1).
• It is not a valuation of any real company. Each buyer is a model company built from published figures for comparable companies (§4.5).
• It produces one central set of results, not a range of random outcomes. Uncertainty is explored by changing inputs and running the model again.

0.4Scope

• There are three cases, each with its own model company: a chocolate maker buying cocoa from Côte d'Ivoire, a coffee roaster buying from Minas Gerais in Brazil, and a brewer buying barley from Castilla y León in Spain (§2.4).
• Each crop has three sites, nine in all. They are real farming areas (§2.4).
• The model runs from 2026 to 2040, plus a value for the years after.
• Company value is modelled at the start of 2027, one year after the programme starts. The companion forestry and water methodologies model it at the start of 2026. Modelling it a year later has two effects. The company's 2026 programme payments fall before the valuation date, so they do not reduce value created, although they are still counted in what you commit. And every later year is discounted for one year less. Both effects make value created higher than modelling it at 2026 would. At Nawa, modelling it at 2026 gives USD 7.3bn created by funding it yourself and USD 7.7bn by buying credits. At 2027 the figures are USD 8.0bn and USD 8.4bn, so the 2026 figures are about 9% less funding it yourself and about 8% less buying credits.
• The company depends on growing conditions for the crop it buys. In the terms of ENCORE, a database of how economic activities depend on nature, these are local climate regulation and soil quality regulation.
• A separate farm-level model uses the same crop data to estimate what the change is worth per hectare at the farm. It is not part of this document, which covers the buyer only.

0.5How to read this

For a short summary, read Part 1 and §7.2.

1.1Two futures for the same landscape

The user picks a crop and a site. The model then builds two complete futures for that landscape. In the first, the company does nothing. We call this future doing nothing. In the second, the company pays for the programme, which is the change in how the farms it buys from are managed. Each future runs through the same chain of steps, from supply and demand to what it means for the company's finances. The difference between the two futures is the value of the programme.

Both futures run on the same model, with the same data and the same assumptions. The only difference between them is whether the company pays for the programme. So any difference between the two answers comes from the programme.

Once it has worked out what the programme is worth, the model turns to the second question. This asks whether buying nature credits gives a better return than funding the work yourself.

1.2What the return is made of

The return has a benefit side and a cost side.

The benefit has three parts.

• Costs avoided. When yields fall in the region the buyer sources from, the world price of the crop rises, and the buyer pays that higher price on everything it buys of the crop. The world price is the same with and without the programme. With the programme, the buyer has a contract with the programme's farmers. When world prices rise, the contract fixes a lower price on the crop this landscape supplies (§3.4).
• Sales kept. When the farms grow less, part of the missing crop cannot be bought elsewhere at any price. The buyer loses that crop and the sales of the finished goods it would have made from it. The programme reduces that loss.
• Cheaper borrowing. Lenders and shareholders see a steadier supply as less risky, so they accept a slightly lower return. That lowers the cost of capital used to value the company, which makes its future cash worth more today. This part counts only in value created. It never appears in extra cash, for the reason given in §1.3.

So costs avoided come from the price contract. The better yields the programme brings help the buyer through the other two parts. The buyer keeps crop that would otherwise be lost, and the sales of the finished goods it becomes, and it borrows more cheaply.

The cost of the programme has two parts. The first is the programme itself. It is the cost of establishing the new practice until it is fully in place in 2030 (called delivery), then maintenance each year, divided by the number of companies paying for it (§5.2). The second is the running costs around it: MRV and verification, coordination, and, where credits are bought, the developer's financing charge.

1.3Two ways to express the return

The model reports two returns. Both divide by the same amount, what you commit: the company's own programme payments from 2026 to 2040, not discounted. They differ in what is divided.

value created         = company value with the programme − company value doing nothing
                        (both modelled at the start of 2027)
extra cash            = Σ free cash flow with the programme − Σ free cash flow doing nothing
                        (2026 to 2040, not discounted)
what you commit       = Σ the company's own programme payments, 2026 to 2040, not discounted
times what you put in = value created ÷ what you commit   (and extra cash ÷ what you commit)

Extra cash is free cash flow with the programme, less free cash flow doing nothing, added up over 2026 to 2040 and not discounted. Value created is company value with the programme, less company value doing nothing, at the start of 2027. Each of them, divided by what you commit, gives a number of times what you put in.

Both returns are reported because they answer different questions.

Extra cash measures the cash the programme produces, after tax and after paying for the programme. It leaves out any change in what the business is worth, so it leaves out cheaper borrowing entirely. Use it to compare the programme with any other project competing for the same budget.

Value created measures what the programme is worth to the company. It includes the effect on operations, the change in working capital and cheaper borrowing. It also includes the terminal value, which is the value of the company's cash flows after 2040, when it still has the more secure supply the programme created. Use it to see how much the programme adds to what the company is worth.

Neither measure is always the larger, and the two can rank options differently. They are built differently. Extra cash adds up 15 years without discounting. Value created discounts, but adds a terminal value and the effects on borrowing. In the worked example in §6.4 they are 4.8 times and 4.4 times what you put in when funding it yourself. At those settings both rank buying credits first. At a higher effectiveness the ranking splits. Buying credits still creates more value, but funding it yourself returns more for each USD committed (§6.4).

Each multiple compares a total over many years with the total committed over 15 years. It is the value created, or the extra cash, for each USD 1 committed. It is not a yearly rate of return, which is why the numbers look large.

1.4Buying credits or funding it yourself

There are two ways to pay for the programme. To fund it yourself, the company pays for the work directly and expenses it. To buy credits, the company buys credits from a developer who does the work. Both routes pay for the same farms and get the same physical protection, so the benefits are the same on both routes. The routes differ only in cost, timing and the points in the table below.

A test checks that the list is complete. Switching every difference off at once makes the two routes equal (§7.4). If an advantage were missing from the table, the test would fail.

Every later step uses one physical input: how much less, or more, the land will grow with and without the programme.

2.1Where the physical numbers come from

The physical numbers come from CropSuite, a published model of how well land and climate suit a crop (Zabel et al., Geoscientific Model Development 18, 1067–1092, 2025). For a given crop and place, CropSuite works out how well that crop can grow. We use CropSuite's suitability score, which we call the yield index because it is scaled to real yields below. CropSuite can also run the same place again with regenerative practices in use, which gives a second result.

Inputs. CropSuite combines climate, soil, terrain and irrigation data. Daily rainfall comes from CHIRPS v2.0 and daily temperature from CHIRTS v1.0. Both come on a 2.5 arcminute grid and cover almost the whole world between 50°S and 50°N. Soil data come from ISRIC SoilGrids at 250 m, in six depth layers from 0–5 cm to 100–200 cm. They cover sand, clay, texture class, organic carbon, pH, electrical conductivity and soil depth. The WISE database adds base saturation, gypsum and sodicity, and the ISRIC Global Soil Salinity Map adds salinity. Slope is worked out from SRTM, a satellite dataset of land height. The extent of irrigation follows Meier et al. (2018).

Process. Statistical methods convert the climate data from a 2.5 arcminute grid to a finer 30 arcsecond grid (about 1 km), in one of two ways. The first uses how temperature changes with height, measured over a moving window and kept within the rates at which rising air cools. The second applies past WorldClim patterns as monthly factors.

The model then works out climate suitability from temperature, rainfall and how much the climate varies. It works out soil suitability from limits set by the soil and the slope. It combines the two under Liebig's law of the minimum. This means it takes the lowest suitability across all the limits, and does not average them.

To judge how much the climate varies, the model tests each possible sowing date over a 20-year period. For each year it checks whether temperature or rainfall crosses a limit during the growing season. The share of years in which one does is how often the crop could fail. As crop failures become more frequent, suitability is reduced along an S-shaped curve, and it falls to zero once the crop would fail in more than one year in four (25%).

To test for more than one harvest a year, the model tries every combination of sowing dates within a year. It allows time for fieldwork between harvest and sowing again. It returns one, two or three possible harvests. Some crops need a cold spell before they flower (vernalisation). The model handles this by counting runs of days cold enough to meet the crop's need.

Outputs. The model gives climate, soil and crop suitability, each on a 0–100 scale. It also gives the best sowing and harvest dates, how many harvests a year are possible, and how often the crop could fail. Last, it gives how much each factor limits the crop, and the one factor that limits it most at that place. Results come on a 30 arcsecond grid, about 1 km² at the equator, as GeoTIFF or NetCDF files.

Limiting factors. Suitability is the lowest score across all the limits, so every place has one factor that holds the crop back most, and the model reports it. In the reference publication, soil is the main limit across tropical regions. Rainfall is what stops maize from being suitable across large dry areas of Africa. This output means the result can be traced to a cause. It also shows whether a given regenerative practice can plausibly ease that factor.

Crop coverage and reproducibility. The reference publication sets up 48 crops. They include maize, rice, wheat, barley, sorghum, millet, cassava, yam, banana, cowpea, sesame, safflower, mango, guava, coffee, cocoa, oil palm and rubber. The authors gave weight to staple and cash crops that matter in their regions but are usually little studied. The three cases in this document use three of those crops. The method is not tied to them. Any crop in the CropSuite set can be used, in any region its climate inputs cover. The user supplies the matching yield index values and a model company for the buyer. That is how the method extends beyond the three cases here.

Turning an index into kilograms

The yield index has no unit. It is tied to real production using SPAM 2020 V2r2 (IFPRI), a global dataset of crop production. The regional average yield is weighted by production. It is total production divided by total harvested area, across the region's grid cells.

actual yield(state) = SPAM regional average yield × [ index(state) ÷ index(2020 reference) ]

The 2020 index on the bottom of the fraction comes from the same climate scenario. So actual yield in 2020, with farming as it is today, equals the SPAM average exactly. Every other year and way of farming is scaled from that figure. An index 10% lower means a yield 10% lower.

ADM1 and ADM2 are the first and second levels of a country's regions. For Spain, the first level is the large region Centro (NUTS-1), which combines Castilla y León with Castilla-La Mancha and Extremadura. Taking Spain at the first level would blend three regions, so barley is taken at the second.

2.2What the programme does on the ground

The programme pays farms to change how they farm. The practice in each case is one that is feasible for that crop in that landscape.

Cocoa and coffee are tree crops that grow well in shade, and cooling from the canopy is the main benefit. Barley is a cereal sown each year in a fairly dry region. Tree shade does not apply there. The main benefits are saving water and building soil carbon. That is also why its programme costs come from a separate body of studies (§5.1).

CropSuite runs each landscape twice: once with farming as it is today, and once with the new practice. The two runs are the same in 2020 and differ from 2030 onwards. The difference between them is the physical effect of the programme. What it costs is in §5.1.

The run with the new practice is optimistic, but within what the evidence supports. It is about the best that excellent management could achieve for that crop in that landscape. It is based on a study-by-study review of field trials and of studies that pool many trials. It is not a promised result for any one field.

2.3Climate

The model uses two climate scenarios, SSP3-7.0 and SSP5-8.5. CropSuite gives the yield index for each place by decade to 2060; the model reads 2020, 2030 and 2040 and draws straight lines between them. No low-emissions scenario was run, so the climate setting has no Favourable level. SSP3-7.0 is Reference and SSP5-8.5 is Adverse (§6.2).

The climate is the same with or without the programme.

The reference publication shows the model over Africa for 1991–2010. The runs used here extend it to future decades and, for coffee and barley, to regions outside Africa, which the near-global coverage of its climate inputs allows.

2.4Sites, and how they were chosen

The barley case depends on one assumption. Evidence on the barley market as a whole says a barley buyer is not much exposed. Barley from one place can replace barley from another. When Spain grows less, imports fill the gap, and a Spanish shortfall does not move European prices. So barley is modelled as a narrower case. The buyer is a brewer that has committed to using barley from Castilla y León. It cannot switch to barley from elsewhere without breaking that commitment.

Each crop has three sites. They are real farming zones within the region CropSuite was run for. The site sets how many hectares the programme covers.

Sourced volume is worked out from the area:

sourced volume = programme hectares × regional yield

A smaller programme covers fewer farms, so it protects less supply. The third site of each crop is the reference site. Its area matches all the volume the buyer sources from that region.

Two points about this design follow. First, CropSuite was run for the region, not for each zone. So all three sites of a crop use the same regional yield figures and the same climate data. The site changes the size of the programme. It does not change the percentage change in yield (§3.1). Telling zones apart would need CropSuite runs for smaller areas.

Second, the share of the buyer's purchases that the model treats as affected by this landscape grows with the site's size, as a fraction of the reference site, up to all of the crop the buyer buys. Without this, the cost of a smaller programme would fall with its area, but costs avoided would not fall with it. A smaller programme would then show a higher return only because of that mismatch. With the scaling, the number of times what you put in is close to the same at every site. A larger site creates more absolute value, but the return for each USD committed stays about the same.

This part works out the two parts of the return that come from day-to-day trading, costs avoided and sales kept.

3.1The physical measure

The change in yield is expressed as one number a year, z:

z = 1 − ( yield index in this year ÷ yield index at the 2020 baseline )

z can be positive or negative. A positive value is a shortfall against the 2020 baseline, and a negative value is a surplus. With the new practice, the yield index can end up above the baseline. For cocoa it more than doubles. The model keeps that surplus, because the contract gives the buyer first call on the extra crop (§3.4).

With the programme, each year's yield change is a mix of the two CropSuite results. The protected share (§3.6) follows the result with the new practice. The rest follows the result with farming as it is today. In the formula, "business as usual" means doing nothing and "fully regenerated" means the new practice on every farm.

z(invested) = z(business as usual) × (1 − protection)
            + z(fully regenerated) × protection

3.2How a shortfall and a surplus differ

A shortfall and a surplus do not act in the same way, so the model treats the yield change in two ways, one for its effect on price and one for its effect on the volume the buyer can get.

z(price)  = MAX(0, z)                    one-sided
z(volume) = MAX(−maxExpansion, z)        symmetric, capped

The price effect counts only shortfalls, because a local surplus does not lower the world price. Growing more cocoa on one set of farms does not make world cocoa cheaper.

The volume effect counts both shortfalls and surpluses, up to a cap. A surplus is useful to the buyer, because it is extra volume from a source it already controls. But it is useful only up to the amount the buyer can take. The cap (maxExpansion) is that amount. The right of first refusal in §3.4 is what gives the buyer access to the surplus.

The model first mixes the two yield results, keeping any surplus, and only then applies the two rules above. Done the other way round, the surplus from the new practice would be dropped before it could make up for the shortfall on farms still farmed as today.

3.3How prices are set

The world price is the same in every future. The buyer's volume from this landscape is a small share of world supply, and its programme does not move the world price. Only the price the buyer pays differs between futures.

world price ratio = ( 1 + elasticity × z(price, business as usual) ) × spike multiplier

The first part is the long-term trend. As yields in this growing region fall, the world price rises. Price elasticity sets how strongly. At 1.5 for cocoa, a 10% shortfall raises the price by 15%. The trend uses the shortfall doing nothing, so it is the same with and without the programme.

The second part is a price spike that we set by hand. It does not come from the crop model. Working out a world price from the model's own yield index would need a model of global supply and demand, which this method does not have. Four patterns of price spikes over 2026–2040 are defined for each crop (Appendix C). The tool offers three of them as the Favourable, Reference and Adverse levels of the price swings setting (§6.2). They are one spike with full recovery, a spike followed by a higher plateau, and two spikes. The fourth, with no spike, is used only for testing. A second setting, normal price, sets the starting price at 0.8, 1.0 or 1.2 times the reference price (USD 4,906/t for cocoa).

3.4The price contract

With the programme, the buyer does not just pay the world price. It has a contract with the farmers in the programme. The contract sets both what the buyer pays and what it can buy.

What the buyer pays. The contract sets a base price and a floor. The base is the world price on the 2026 contract date. The floor is 110% of the base. Two cases follow.

If the market price falls below the base, the buyer still pays the floor. The farmer is shielded from the fall and gets more than the market would pay. At cocoa's default settings the base is USD 5,740/t and the floor is USD 6,314/t. So with the market at USD 4,000/t, the farmer still gets USD 6,314/t.

If the market price rises above the base, the buyer pays the base plus an agreed share of the rise. The farmer gets part of the rise, and the buyer pays less than the market. At the same settings, with the two-spike price pattern, a market price of USD 12,783/t in 2035 gives a paid price of USD 9,261/t. That is USD 5,740 plus half of the USD 7,043 excess.

paid = coverage × MAX( floor, base + share × MAX(0, world − base) )
     + (1 − coverage) × world

Contract coverage. The contract is signed with the programme's farmers, so it can cover only the tonnes this landscape supplies. The buyer buys the rest of its crop from other regions at the world price, and the contract does not cover it. So coverage is worked out, not set by hand. It is this landscape's tonnes divided by all the crop the buyer buys. The buyer's total purchases are converted into tonnes using the reference price, not the scenario price, so the share does not change when the price setting changes. For cocoa that is 150,000 t of 532,002 t, or 28.2%. Coffee is 48.7% and barley 7.3%. The share is the same at every site, because a smaller site supplies fewer tonnes and the model treats a correspondingly smaller part of the buyer's purchases as affected (§2.4).

Coverage is not set at 100%. That would give the programme's price terms to crop bought from farmers outside the programme, and add a large benefit the programme does not earn. For coffee the effect would be the opposite. The floor is above the world price in most years, so a contract on all the crop the buyer buys would cost more than it saves.

What the buyer can buy. The contract also gives the buyer a right of first refusal over extra crop grown at the site. This is what makes the deal worthwhile for the buyer. In a poor year, the buyer keeps access to supply from farms it has paid to protect. In a good year, when the programme lifts the site above its baseline, the buyer has first call on the extra volume. Without that right, the buyer pays for the improvement and someone else buys the extra crop, so the gain goes to them. The right of first refusal is what lets the buyer benefit from the surplus in the volume effect (§3.2).

The base is not an average of recent market prices, because such a base lags the market. Once a spike passes, the average stays above the market, and because the floor is set from that average, the buyer would keep paying more than the market for several years. A base fixed at the contract date does not do this.

3.5The two operating channels

Revenue effect = origin volume × z(volume) × (1 − covered)
                 × reference price × downstream conversion

Cost effect    = exposure × commodity-cost share of COGS × base COGS
                 × ( paid price ratio − 1 )

In the formulas, origin volume is the tonnes from this landscape, and covered is the replaceable share described below.

The part of a shortfall that cannot be replaced from other sources is truly lost. It is valued as finished goods, not as raw crop, through a multiple that converts the value of the crop into the value of the finished goods (the downstream conversion). A tonne of cocoa the buyer cannot source does not cost it USD 4,906. It costs the lost margin on the chocolate that tonne would have become.

The replaceable part is bought from other sources at the market price. Paying that price raises the cost of goods sold (COGS), in proportion to how much of the buyer's costs this crop makes up.

At the reference site, the model assumes a change in the world price affects all of the crop the buyer buys, because it has one world price per crop. This share is called exposure. It is a modelling convention, not an estimate from data. It grows with site size as described in §2.4. Exposure and contract coverage are different things. Exposure is the part of the buyer's purchases that the world price affects. Contract coverage is the part bought under the contract, which is this landscape's tonnes (§3.4).

3.6Protection

The buyer does not get the full physical effect of the new practice at once. Three things reduce it. The practice takes four years to establish, the programme may not include every farm the buyer buys from, and not all of the work succeeds.

protection(y) = ramp(y) × secured share × effectiveness
ramp(y)       = u² × (3 − 2u),   u = (y − 2026) ÷ (2030 − 2026), held between 0 and 1

Protection builds up over four years, from the contract date (2026) to the date the new practice is fully established (2030, called delivery). It follows an S-shaped curve, slow at first and fastest in the middle, with values of 0, 16%, 50%, 84% and then 100% from 2030. In the formula, ramp is this build-up. The secured share is the part of the buyer's purchases from this landscape that comes from farms in the programme. At the defaults it is 100%. Every farm the buyer buys from here is in the programme. Effectiveness is the share of the work that succeeds. It is set by the programme effectiveness setting: 0.85 at Favourable, 0.60 at Reference and 0.35 at Adverse (§6.2).

Effectiveness scales the physical benefit on both routes. It also scales what a credit buyer pays (§5.4) and the reward from banks and investors (§4.6).

3.7What the buyer is buying

The buyer pays so that less crop is lost on the farms it buys from. It sees the benefit as fewer lost sales and cheaper borrowing. With it comes a contract that sets the price it pays for this landscape's crop, which lowers its costs when world prices rise, and gives it first call on any extra crop the programme grows (§3.4).

Each credit stands for work on the farms that has been done and independently checked. It gives the buyer no right to any tonnes of crop. That right comes from the contract. Without the contract, the buyer would pay for the improvement and another buyer could take the extra crop.

The credit pays for work that would not otherwise happen (it is "additional"). The new practice pays for itself at the farm over time, but getting it started costs money. Farmers earn nothing extra from the practice until it is established, so they have no income to borrow against. The credit pays for that start.

This part sets out the company's accounts. They add two effects to the return. One is the cash tied up in holding stock of the crop (working capital). The other is cheaper borrowing.

4.1Why a full set of financial statements

The model treats the buyer as a company with a full income statement, balance sheet and cash flow statement. It values the company once for each of the three options.

The reason is that a nature-credit programme acts on a company in three separate ways, which show up in different places. It changes operating margin and revenue. It changes the working capital needed to hold stock. It changes the rate at which the company's future cash is discounted. Only a full set of linked statements records each effect once, in the right place. A balance sheet that must balance also checks that no effect has been counted twice or left out.

The other way to build it is to work out each effect separately and add them up. That approach can go wrong where the separate calculations overlap. The linked model avoids two failures in particular. First, a benefit both routes get could end up counted for only one of them, because nothing forces both routes to use the same accounts. Second, a ratio such as the downstream conversion could be applied to a revenue or cost figure it was not estimated from, because nothing ties the ratio and the figure it multiplies to the same source. A single linked model prevents both.

The model also gives a result buyers are used to, which is what the company is worth under each of the three options.

4.2The three futures

How the programme is accounted for is fixed. The user cannot choose it. Funding it yourself is expensed, so it gets its tax relief at once. Credits are recorded as an asset and written off through the income statement, which delays the tax relief. Recording credits as an asset assumes the buyer can show that it controls the future benefit they bring. For a buyer that sources from the landscape, the strongest basis for this is the contract, which gives it first call on the extra crop the programme grows (§3.4). A valuation built on cash does not depend on how the cash is accounted for, except through tax. So any difference the model reports between buying credits and funding it yourself comes from the size of the costs, the timing of tax, or one of the differences named in §1.4.

4.3Where each effect is recorded

Each effect is recorded in one place. The two operating effects, costs avoided and sales kept, appear in the income statement. The cost of capital is the discount rate used to value the company. Interest is charged on the debt balance, so anything that changes debt also changes interest. Working capital is the value of stock on the balance sheet. When stock costs more to hold, the company borrows more to pay for it, and free cash flow falls in the year the stock rises.

In the cocoa case, the 2023–2025 price rise hurt the cocoa industry mainly through working capital. Companies had to put up extra cash against their futures contracts (margin calls) and pay more to hold stock. As bean prices tripled, the working capital needed to hold the same physical stock tripled too. At least one major processor's net debt about tripled as a result. That is a balance-sheet effect. It does not show in a profit-and-loss calculation, or in a cost-of-capital calculation. A model with all three statements captures it directly.

4.4Circularity

Holding more stock means more debt, more debt means more interest, more interest means less cash, and less cash means more debt again. To avoid this loop, the model charges each year's interest on the debt at the end of the year before.

interest(t) = cost of debt × net debt(t − 1)

The workbook can therefore calculate each year once, in order.

4.5The model company

Each case uses a model company: a stylised company set to match real comparable companies. It matches them on size, margins, working capital, debt, cost of debt and tax rate. Full parameters are in Appendix B.

Two values of the model company are worked out, not set by hand, so that they stay consistent with the company's revenue and earnings when those change. EBITDA is earnings before interest, tax, depreciation and amortisation:

fixed assets     = revenue × (fixed assets ÷ revenue) ratio
company value    = base EBITDA × (company value ÷ EBITDA) multiple

For cocoa both come from a comparable company's published figures. For coffee and barley the ratios are worked back from the published figures of each case's comparable company, so the method is the same in every case.

The multiples of company value to EBITDA that result are 12.5× for cocoa, 12.07× for coffee and 8.55× for barley. Cocoa and coffee are within 4% of each other. They came from separate comparable companies and were not tuned to agree. Barley is well below them. That is the direction you would expect for a private regional brewer, but barley's figures come from the weakest comparable-company data of the three (§7.5).

4.6Valuation and cost of capital

doing nothing:       w = base rate
fund it yourself:    w = base rate − cut × benefit scale
buy credits:         w = base rate − ( cut + duration premium + verification premium
                                        + cheaper-debt benefit ) × benefit scale

benefit scale = secured share × effectiveness

In the formulas, w is the cost of capital and "cut" is the cut in it for a steadier supply. The three extra terms on the credit line are the duration premium, the verification premium and the cheaper-debt benefit (§1.4).

The cut in the cost of capital is scaled by how much of the supply the programme secures and how much of the work succeeds. A programme that secures nothing earns no cut.

How much benefit lenders and investors give is set by the reward from banks and investors setting (§6.2). It scales the cut in the cost of capital and the three credit premiums together, to 100%, 50% or 0% of their base values. The base values are 0.25 percentage points for the cut in the cost of capital, 0.03 for the duration premium, 0.01 for the verification premium and 0.05 for the cheaper-debt benefit.

The cheaper-debt benefit lowers the cost of capital only in proportion to the company's share of debt, after tax:

cheaper-debt benefit in the cost of capital = cost-of-debt cut × (1 − tax) × net debt ÷ company value

Counting a saving on the cost of debt as if it were a saving on the whole cost of capital would overstate it by 6 to 9 times, depending on the crop.

Company value is what the company is worth to lenders and shareholders: its future cash in today's money. In a given year, it is the present value of free cash flow from that year on, including that year's own flow. A terminal value at 2040, from a standard formula that assumes steady growth (the Gordon growth formula), covers the years after. In the formulas, FCF is free cash flow, w is the cost of capital and g is the growth rate after 2040:

EV(2040) = ( FCF + FCF × (1 + g) ÷ (w − g) ) ÷ (1 + w)
EV(year) = ( FCF(year) + EV(year + 1) ) ÷ (1 + w)

Company value is modelled at the start of each year, so the first year's cash flow is included in the result. This matters because funding it yourself puts its largest payment early.

This part builds the cost side of the return.

5.1Programme cost

Cost depends on the size of the site:

establishment = hectares × establishment rate (USD/ha)
annual cost   = hectares × annual rate (USD/ha/yr)

Rates come from a table of three tiers, split by practice, not by crop. Cocoa and coffee share one body of evidence for agroforestry. Barley's no-till practice is a different kind of work. Its costs come from separate studies, and it is much cheaper.

The three tiers are the three levels of the programme cost setting (§6.2). Sources are in §7.3. The evidence is uneven. The Favourable tier has no citation of its own, and the Adverse tier is based on a single documented programme.

5.2Timing and who pays

Establishment runs from the contract date, 2026, to the delivery date, 2030. Maintenance runs from delivery onward. The buyer pays 1/N of the programme, where N is the number of companies paying for it, including the buyer. At the defaults N is 1, so the buyer pays alone.

The routes differ in when the buyer pays. Under a credit contract the developer can borrow against the buyer's promise to buy delivered credits, so the buyer pays from delivery onward. An agreement to share costs between companies is not something a lender will accept as security, so a buyer funding the work itself pays for establishment up front.

5.3The developer's financing charge

The developer borrows to pay for establishment and is paid back through the credit price after delivery. The developer's cost of borrowing the money until delivery is included in the credit price.

carried establishment = Σ over 2026–2029 of ( establishment(y) + MRV(y) ) × (1 + r)^(2030 − y)
yearly recovery       = carried establishment × r ÷ (1 − (1 + r)^−T)
credit purchase(y)    = ( maintenance(y) ÷ N  +  MRV(y) ÷ N  +  yearly recovery ÷ N, for T years
                          from 2030 ) × payment factor (§5.4)

In the formula, carried establishment is the establishment and MRV and verification spending before 2030, with interest at the developer's rate to 2030. r is the developer's rate, 7%, and T is the period over which the developer recovers its cost, 10 years. Both are judgements. At Nawa the programme's USD 467m of establishment, with its MRV and verification, comes to about USD 560m by 2030, once the developer's interest is added, and is recovered at about USD 80m a year from 2030 to 2039.

5.4Risk transfer

A credit buyer pays only for work that is delivered, which is 60% at Reference. When funding it yourself, the company commits the spending whatever is delivered. The developer bears the risk on the other 40% and passes half of it back to the buyer, raising the price of each delivered credit by 20% (half of 40%). So the buyer pays 72% of the full cost. In the formula, the share of risk repriced is the half the developer passes back.

payment factor = effectiveness × ( 1 + (1 − effectiveness) × share of risk repriced )
               = 0.60 × ( 1 + 0.40 × 50% ) = 0.72

Effectiveness also scales the physical benefit (§3.6), so a less effective programme lowers the benefit equally on both routes. Credits gain only by paying less. They receive no more benefit than funding it yourself.

The table shows risk transfer on its own. The cost-of-capital benefit is held fixed, so that it does not shrink as effectiveness falls, and the only difference left is what each route pays. Cocoa, Nawa, every other setting at Reference.

Risk transfer works like insurance. When the programme delivers in full, buying credits is only slightly ahead, through the cost-of-capital premiums (§1.4). As delivery fails, the gap widens, because the credit buyer stops paying for work that does not happen. From 100% to 40% effectiveness, the credit buyer pays about USD 1.0bn less, and buying credits gains about USD 0.6bn of value created on funding it yourself. That offsets about 39% of the USD 1.5bn benefit lost on both routes.

The model understates the value of passing risk to the developer. It produces one central set of results, so it counts only the payment avoided on average. Passing on the risk also protects the buyer from results worse than average, which one central set of results cannot show. So the figure is a floor.

5.5MRV, verification and coordination

MRV and verification means measuring, reporting and verifying the work. It costs USD 3 a hectare a year, on the company's share of the hectares, every year from 2026 to 2040. It is paid on both routes.

If the company funds the work itself, it pays MRV and verification each year as it is spent. If it buys credits, the developer pays these costs and recovers them in the credit price (§5.3). Costs before 2030 are added to establishment, and costs from 2030 are added to the yearly price. They are then treated like the rest of the price. MRV and verification are not counted as an advantage of either route.

Coordination is the one-off cost of organising the arrangement: USD 0.4m in 2026. Only the fund-it-yourself route pays it, and it is not shared. USD 0.4m is a low estimate. Organising several companies to fund one programme needs contracts, governance and reporting, so the real cost is likely to be much higher, and part of it recurs each year. A higher cost would favour buying credits, because only the fund-it-yourself route pays it.

6.1How the tool is organised

Each crop's case reads left to right through ten steps, from the landscape and the programme to what the company is worth. Each step shows the figure doing nothing and with the programme.

The tool separates what the buyer decides from the conditions the decision is tested against.

Choices are decisions the buyer makes. Each appears at the step where it is needed: crop, site, number of co-funders, the buyer's share of price rises, and how many years a credit is written off over. Whether to buy credits or fund it yourself is not a choice. Both routes are always calculated.

Scenario settings are conditions the buyer does not control. There are six, each set to Favourable, Reference or Adverse (§6.2). They are in one panel that stays in view at every step, next to a chart of the selected settings. The tool opens with every setting at Reference.

Other assumptions are three inputs no setting drives: price elasticity, the base cost of capital, and the MRV and verification cost. They appear beneath the settings. Two of them move the result as much as most settings, so they are shown rather than hidden.

Fixed parameters are every other input. They are not shown.

Each setting has a drawer showing the inputs it drives. Overriding one marks that setting Custom, so the chart never shows a level the inputs no longer match. Changing an other assumption never marks a setting Custom.

6.2Scenario settings

Favourable is good for the buyer, Adverse is bad, and Reference is the central case that the others are compared with. Three settings describe the world; three describe how the programme performs and how it is rewarded.

Climate has no Favourable level. No low-emissions scenario was run, so SSP3-7.0 is Reference and SSP5-8.5 is Adverse. The Favourable level is left empty rather than filled with a high-emissions scenario presented as the good case.

Price swings. The pattern with no spike is used only for testing and is not offered. The three patterns offered are ranked Favourable to Adverse by how much each lowers the cocoa buyer's company value doing nothing. For barley the "spike then plateau" and "two spikes" patterns peak at the same height, and the plateau does slightly more damage. So for barley the Reference pattern does slightly more damage than the Adverse one. The cocoa ranking is kept for all crops so each level means the same thing.

Normal price. Favourable is 0.8 times today's price, Reference is today's price and Adverse is 1.2 times. For coffee and barley the modelled regions grow more to 2040, so there is little shortfall for a price rise to act on, and a higher price changes company value doing nothing only a little.

The record rule. Each crop's price spike patterns are scaled so that the most severe one (two spikes), at the Reference normal price, peaks at the highest price the crop has actually reached. Cocoa's pattern meets this without scaling. Coffee's unscaled pattern would peak 11% above the arabica record, so it is scaled down. Barley's would peak well below the 2022 European record, so it is scaled up (Appendix C). At an Adverse normal price the two-spike pattern runs about 20% above the record for every crop.

Programme effectiveness. 0.85 represents a well-run programme. All published figures are well below it (§7.3), so it is the Favourable level, not the central one.

Reward from banks and investors. The cut in the cost of capital (0.25 percentage points), the duration premium (0.03), the verification premium (0.01) and the cheaper-debt benefit (0.05) move together. At Adverse all four are zero. The green-bond evidence points to a premium that is very small rather than absent, but zero shows clearly what happens with no reward.

6.3What each setting moves

Each row moves one setting from Favourable to Adverse with every other setting at Reference, at each crop's reference site. Value created, modelled at the start of 2027.

Cocoa, Nawa

Coffee, Sul de Minas

Barley, Tierra de Campos

Figures from the model, with every other setting at Reference.

What the tables show:

• Programme effectiveness moves value created most for coffee and barley. For cocoa the normal price moves it slightly more, by USD 2.9bn against USD 2.4bn.
• Programme cost moves value created less than it moves the multiple, because it is also the denominator. For cocoa, what you commit runs from USD 396m at Favourable to USD 4.0bn at Adverse, so value created for each USD committed falls from more than 20 times to 1.6 times.
• Buying credits creates more value than funding it yourself on every row. But at a Favourable programme effectiveness the ranking per USD committed splits. Buying credits needs more cash (USD 1.8bn against USD 1.7bn for cocoa), because at 85% more of the work is delivered and paid for, and the credit price also includes the developer's financing and risk charges. So funding it yourself returns more for each USD committed (4.9 times what you put in, against 4.6).
• Favourable price settings mean lower and calmer prices. That is good for the buyer but makes the programme worth less. Favourable programme settings make it worth more. The two pull in opposite directions, so setting everything to Favourable does not give the highest value created for cocoa or coffee.
• An Adverse climate lowers value created for every crop.

• Two inputs that no setting drives also matter. Moving the base cost of capital from 7% to 9% moves cocoa's value created on the fund-it-yourself route from USD 9.8bn to USD 6.8bn, as much as the largest setting. Moving price elasticity from 1.0 to 2.5 moves it from USD 7.7bn to USD 8.7bn.

6.4Worked example

Cocoa, Côte d'Ivoire, the Nawa reference site, every setting at Reference, every choice and other assumption at its default.

The result, USD m

At these settings buying credits leads on every measure. The reason is effectiveness. Credits are paid for on delivery, so at 60% the buyer pays for fewer credits and needs USD 237m less cash, while the 40% of the work that fails reduces the benefit equally on both routes.

At a Favourable programme effectiveness of 85% the ranking splits. Buying credits still creates more value (USD 8,389m against USD 8,271m), but funding it yourself returns more for each USD committed (4.9 times what you put in, against 4.6), because at this level buying credits needs more cash: USD 1,833m against USD 1,681m. Which route is better depends on whether the company cares more about what it is worth or about the return on its cash. Reporting one measure alone would hide the trade-off.

Site scale changes the size of the result more than the return for each USD committed. With every setting at Reference, funding it yourself creates USD 2.3bn at La Mé (5.5 times what you put in), USD 4.0bn at Indénié (5.1 times) and USD 8.0bn at Nawa (4.8 times).

7.1Confidence labels

7.2How the return is made up, with confidence

The operating benefits are based on evidence. Cheaper borrowing is based on judgement. That is why both measures are reported. Extra cash leaves the judgement-heavy part out entirely.

7.3Sources

Biophysical model and data

• Zabel, F. et al. (2025). CropSuite v1.0: a comprehensive open-source crop suitability model considering climate variability for climate impact assessment. Geoscientific Model Development 18, 1067–1092. doi:10.5194/gmd-18-1067-2025
• IFPRI (2024). Spatial Production Allocation Model (SPAM) 2020 V2r2 Global Data. Harvard Dataverse. doi:10.7910/DVN/SWPENT. Yield and harvested-area layers, all-technology.
• Funk, C. et al. Climate Hazards Group InfraRed Precipitation with Stations (CHIRPS) v2.0, and Climate Hazards Center InfraRed Temperature with Stations (CHIRTS) v1.0.
• ISRIC SoilGrids (250 m); ISRIC WISE database; ISRIC Global Soil Salinity Map.
• Meier, J., Zabel, F. & Mauser, W. (2018). A global approach to estimate irrigated areas. Hydrology and Earth System Sciences 22, 1119–1133.
• FAO. FAOSTAT Producer Prices, element 5532.
• US Bureau of Labor Statistics. Consumer Price Index for All Urban Consumers (CPI-U).

Nature outcome (Part 8)

• Leutner, B. (2025). The Ecosystem Integrity Index: a comprehensive, globally applicable, and unified metric for ecosystem health. Technical whitepaper, The Landbanking Group, Munich, v1.0. Licensed CC BY-SA 4.0.
• Hill, S. L. L., Harrison, M. L. K., Maney, C. et al. (2022). The Ecosystem Integrity Index: a novel measure of terrestrial ecosystem integrity. UNEP-WCMC. doi:10.1101/2022.08.21.504707
• Theobald, D. M., Kennedy, C. M., Oakleaf, J. R. et al. (2025). Global Human Modification v3. Scientific Data. doi:10.1038/s41597-025-04892-2
• Gassert, F., Mazzarello, J. & Hyde, S. (2022). Global 100 m Projections of Biodiversity Intactness for the years 2017–2020. Technical white paper, Impact Observatory.
• Hudson, L. N., Newbold, T., Contu, S. et al. (2014). The PREDICTS database: a global database of how local terrestrial biodiversity responds to human impacts. Ecology and Evolution 4, 4701–4732.
• Copernicus Land Monitoring Service. Gross Dry Matter Productivity. European Union. doi:10.2909/b54be6e3-5962-4e9a-a443-9ff7ea943dfe

Price elasticity

• Tothmihaly, A. (2018). How low is the price elasticity in the global cocoa market? Long-run cocoa demand elasticity −0.34 and supply elasticity 0.57. Converted to this model's measure with a simple one-market model, these imply a value nearer 2.5–3 than the 1.5 used. So the model is cautious on this input.

Commodity prices

• ICE cocoa futures. About USD 2,400/t before the spike; USD 12,931/t at the peak (December 2024); USD 4,924/t (November 2025); USD 2,846/t at the low (April 2026); about USD 5,740/t (August 2026).
• Arabica coffee. Record USD 4.41/lb (February 2025); USD 2.39/lb at the low (June 2026); about USD 3.28/lb (August 2026).
• Tridge. European malting barley, 2025: Spain €234/t, Germany €215/t, Netherlands €187.50/t.
• European Commission, AGRI-food data portal, weekly cereal prices. Spain national average malting barley €388.65/t, week of 16–22 May 2022 (Burgos €389.28/t), the peak of the 2022 price event. Used as barley's price record under the record rule (§6.2).
• International Monetary Fund, Primary Commodity Prices, global price of barley (FRED series PBARLUSDM). All-time high USD 263/t in April 2022, against USD 122/t in July 2026. Supports the timing of barley's record.

Programme cost

• Project Drawdown, Multistrata Agroforestry (via Carbon Brief). Establishment above USD 1,300/ha.
• To'ak Chocolate. Ecuador cocoa agroforestry programme: USD 2,900/ha to set up, plus USD 900/ha/yr over 5 years.
• Winrock International. Financing Smallholder Cocoa Rehabilitation in Ghana. USD 333–5,000/ha/yr for rehabilitation and restoration. This is more intensive work than shade agroforestry alone. So it is used only as an upper reference, not as a comparable rate.
• SARE and DTN Progressive Farmer. No-till losses in the year of change of USD 11.50–39 an acre.
• Mighty Earth and Voice Network. Agroforestry in the Cocoa Sector.

Effectiveness

• In Côte d'Ivoire, under 2% of trees handed out in campaigns with no follow-up have been recorded as surviving. Programmes that check the work and train farmers do much better, but no figure has been published.
• Probst, B. et al. Systematic assessment of the achieved emission reductions of carbon crediting projects. Nature Communications. Across 2,346 projects assessed, under 16% of issued credits stand for real reductions. Used as the closest measured comparison for whether credits deliver.

Cost of capital

• Copenhagen Business School, Nordic ESG Lab. The green bond premium fell from about 0.06 percentage points (2015) to too small to measure by 2020. The premium on government green bonds in advanced economies was about 0.02 percentage points, 2014–2023.
• Harvard Business School Working Paper 23-027. Borrowers on sustainability-linked loans get no significant price cut up front. Lower rates follow the borrower's reputation, not a green premium.

Comparable companies

• The Hershey Company. Net debt/EBITDA 1.90× (June 2026); fixed assets/revenue 30.7%; company value/EBITDA about 12.5×.
• JDE Peet's N.V. Published revenue, margins, inventory days, debt and company value.
• Heineken N.V. Net debt/EBITDA 2.3×, against a stated target below 2.5×.
• Carlsberg A/S. Net debt/EBITDA above 3.0×.
• Mahou San Miguel S.A. Published revenue and EBITDA margin. Gross margin, inventory days and debt are sector estimates, because the company is private.
• Barry Callebaut AG. How working capital and net debt moved through the 2023–25 cocoa price shock.

7.4Validation

Each rebuild is checked against the six tests below.

The decomposition test. Switching off every difference between buying credits and funding it yourself at once brings the gap between the two routes to exactly 0.00. This shows that the list in §1.4 is complete.

The balance check. Assets less financing must equal zero in every year, every scenario and every case. The worst value seen across all combinations of case, scenario and site is 1.2 × 10⁻¹¹. That is rounding noise in the computer's arithmetic.

Cross-case consistency. Every crop, climate scenario and site is calculated in the same workbook by changing the selections, with no separate copy for any case.

The zero-effect test. The programme is set to change nothing: yield with the programme equals yield without it, and the reward from banks and investors is set to none. Value created must then be zero or below on both routes. With the price contract switched off, it passes for all three crops: at each reference site, funding it yourself gives −USD 1.2bn for cocoa, −USD 278m for coffee and −USD 6.8m for barley, and buying credits −USD 906m, −USD 210m and −USD 4.9m. With the contract on, coffee still passes, but cocoa creates +USD 558m funding it yourself and +USD 845m buying credits, and barley +USD 2.1m and +USD 3.9m. This is because the contract improves the price the buyer pays on this landscape's crop whether or not the programme works (§3.4, §7.5).

The contract-coverage test. At every crop, climate, site and price level, the contracted share of all the crop the buyer buys must equal this landscape's tonnes over that total, at the reference price. It must not change with the price scenario. It passes in all 54 combinations.

The fixed test configurations. A rebuild must reproduce 45 recorded configurations: the six crop and climate cases with every setting at Reference, each setting moved alone to Favourable and to Adverse, and every setting at Favourable and at Adverse together (Appendix C). Each must reproduce its recorded figures.

7.5Known limitations

Ranked by how much each would move the answer.

1. The cut in the cost of capital has no source. No published figure measures a cost-of-capital benefit from investing in a more resilient supply chain. Because it is scaled by the protection actually delivered, it moves cocoa's value created by only about 6% between full and no reward (§6.3). But together with the credit premiums, it makes up most of the difference between buying credits and funding it yourself. That is why it is shown as the reward from banks and investors setting, which goes down to zero.
2. Price elasticity drives the result doing nothing more than any other input, and it is sourced for cocoa only. Coffee's is set by analogy with cocoa, and barley's has no source.
3. The credit premiums are judgements, and the advantage of buying credits depends on them. The available studies would support figures close to zero. The model uses a middle value.
4. An effectiveness of 0.85 represents a well-run programme, not the expected case. All published figures are well below it. That is why it is the Favourable level of the programme effectiveness setting, and 0.60 is Reference. The Reference and Adverse levels are judgements, not measurements.
5. Barley's model company has the weakest data of the three. Mahou is private. Gross margin in particular is a sector estimate, and it sets what a lost tonne costs the brewer.
6. The downstream conversion for coffee and barley has no source. Coffee's is an assumption, not worked out from evidence on retail mark-ups, which measure a different part of the chain. Barley's physical ratio of malt to beer is sourced, but the value multiple is not.
7. All three sites of a crop use the same regional yield figures (§2.4). The site changes the size of the programme. It does not change the percentage change in yield.
8. The CropSuite result with the new practice is optimistic (§2.2). It is taken from the source model.
9. Contract coverage uses the landscape's starting tonnes in every year. Where the landscape delivers less than that, coverage is slightly too high. The contract is meant for farmers in the programme, but the model does not make it depend on the programme. With the contract switched off, cocoa's value created by funding it yourself, with every setting at Reference, is USD 6.3bn instead of USD 8.0bn.

10. The nature outcome is described, not computed. The ecosystem-integrity values in Part 8 are estimated by judgement from the published method and what each landscape is like. They are not worked out from data, and they have a lower confidence label than any other figure in the tool. They affect no result. The limits are set out at §8.5.

7.6What is not modelled

What the model leaves out, why, and which way each would move the answer.

8.1What this part covers, and what it does not

Parts 1 to 7 describe a financial method. Nothing in them depends on nature being measured. The model uses the yield index from CropSuite, and the financial results are worked out from it.

This part covers the separate question of what the programme does to the land itself. It is descriptive. No figure in it is used in the valuation, so changing it does not change any result.

Healthier nature on the farms, including more biodiversity, supports better yields. The buyer benefits through that yield, so the model values the yield and does not add a separate value for biodiversity.

8.2The index

The tool reports ecosystem integrity, from the Ecosystem Integrity Index.

The index scores a place by how it compares with what it could naturally support, using three measures, called pillars.

The compositional pillar is measured against the community that was originally present, not against a count of what is there now. A species arriving because the land has been disturbed does not raise it; a native species lost does lower it.

Sources are listed at §7.3.

8.3The limiting-factor rule

The index takes the weakest of the three pillars, not their average. It then lowers the result further where the other two are also poor.

index = weakest pillar × ( a + b − a × b )
a, b  = the other two pillars, each from 0 to 1

The bracket is close to 1 when either of the other two pillars is close to intact, and it falls as both fall. Condition and change are worked out from this index (§8.4).

The reasoning is that a landscape which is productive and well connected but has lost its native community is not a healthy one, and an average would hide that. A severe failure on any one pillar limits the overall score.

This is the same rule the crop model in Part 2 uses. In CropSuite the scarcest input caps attainable yield (§2.1). Here the weakest pillar caps condition.

One result of this rule is that a programme that lifts a pillar which is not the weakest changes the score less than it changes the land. Shade agroforestry raises productivity a good deal on a cocoa landscape. It also brings back part of the native community that a full-sun plantation shuts out. What holds these landscapes back is structure, because they remain farmed. So the score understates the improvement. The real gain may be larger than the score shows.

The rule also shows what would improve the score most. On a landscape held back by structure, what raises the score is land taken out of farming as habitat. Examples are strips along the watercourses, field margins, and patches kept between the blocks. No change inside the fields does that. This is a point about how the programme is designed.

8.4How condition and change are described

The tool reports condition and change in words, not numbers, because the underlying values are reasoned rather than computed (§8.5). Each word matches a fixed range of values, so that it means the same thing on every landscape.

Condition, on the index value: Heavily modified (below 0.20) · Substantially modified (0.20–0.40) · Partly intact (0.40–0.60) · Largely intact (above 0.60).

Change, on the share of the gap to natural potential that the programme closes: Slight (under 3%) · Modest (3–7%) · Moderate (7–15%) · Marked (15–30%) · Major (above 30%).

Change is expressed as a share of the remaining gap, not as a percentage of the index value. A percentage of the value would reward degradation. The same action on a degraded landscape would score several times higher than on an intact one, because its starting point is smaller. Share of the gap treats every landscape the same, whatever its starting point.

8.5What this cannot tell you

• The index is not computed for any landscape in the tool. The pillar values are estimated by judgement from the published method and what each landscape is like. Nothing was measured on the ground or derived from satellite imagery for these sites. Every figure has an indicative marker. This is a weaker claim than the judgement marker used elsewhere in this document (§7.1). A Judgement label marks a stated value for an input the model uses in its calculations. The ecosystem-integrity values are not used in any calculation and were not computed; they are informed estimates.
• The natural baseline is the best-protected areas that exist today, not a pre-industrial state. So the index measures how far a place is from the best-protected land left today, not from its original state.
• There is no freshwater pillar. The index does not cover rivers, groundwater, river flows or wetlands.
• The index says nothing about threatened species or extinction risk. A landscape can improve on this measure while a species on it continues to decline. Covering that requires a separate species-risk layer, which the tool does not have.
• The structural and compositional pillars both draw on human-pressure data and so are not fully independent. The index's authors identify this as an open problem in their own account of the method.
• The index is terrestrial, and it is a measure of ecosystem condition. It should not be described as a biodiversity metric: species composition is one of its three pillars, not the whole of it.

Appendix ACalculation order

The model works through these steps in order. The code uses different names for some terms. In it, delivery is effectiveness, carry is the developer's financing charge, Direct is funding it yourself and Credit is buying credits. S1 and bau are doing nothing, and S2 is with the programme. Invest value is value created, cash out is what you commit, and WACC is the cost of capital.

 0.  scenarios       six Favourable/Reference/Adverse selectors resolve climate pathway, spike profile,
                     reference-price multiple, cost tier, delivery rate and the capital-recognition
                     scale on ΔWACC and the credit premiums (§6.2)
 1.  idx_ref(y)      interpolate 2020 baseline → 2030 ref → 2040 ref
 2.  idx_regen(y)    regen2030 before 2030, then interpolate → regen2040
 3.  ramp(y)         smoothstep u²(3−2u), u = (y − contract start) ÷ (delivery − contract start)
 4.  protection(y)   ramp × secured × delivery
 5.  z_bau(y)        1 − idx_ref ÷ baseline
 6.  z_inv(y)        z_bau × (1 − protection) + (1 − idx_regen ÷ baseline) × protection
 7.  z_price         MAX(0, z)
 8.  z_volume        MAX(−maxExpansion, z)
 9.  world_trend(y)  1 + elasticity × z_price(bau)
10.  world(y)        world_trend × spike multiplier(y)
11.  base            world(2026), frozen
12.  floor           MAX(absolute floor ÷ reference price, floor multiple × base)
13.  paid(y)         S1: world.  S2: coverage × MAX(floor, base + share × MAX(0, world − base))
                         + (1 − coverage) × world
14.  hectares        site library, selected by crop × site
15.  volume          hectares × regional yield
16.  exposure        MIN(1, case exposure × hectares ÷ reference-site hectares)
17.  Revenue         base revenue − volume × z_volume × (1 − covered) × ref price × conversion ÷ 1e6
18.  COGS            base COGS − (base revenue − Revenue) × (1 − gross margin)
                         + exposure × commodity share × base COGS × (paid − 1)
19.  Programme       establishment = hectares × estab rate ÷ 1e6, spread contract→delivery
                     maintenance   = hectares × annual rate ÷ 1e6, from delivery
20.  Payments        Direct: programme ÷ N as incurred, plus MRV ÷ N each year from the contract date,
                             plus coordination at contract date
                     Credit: (maintenance ÷ N + MRV ÷ N + annuitised carry) × credit payment factor,
                             from delivery
21.  Statements      EBITDA → amortisation → EBIT → interest (one-year lag) → tax → net income
                     Inventory = base inventory × (1 + exposure × (paid − 1))
                     Net debt, equity, balance check = 0
22.  FCF             EBIT × (1 − tax) + amortisation − Δinventory − capex
23.  WACC            §4.6
24.  EV              start-of-year, Gordon terminal at 2040
25.  Invest value    EV(invested, 2027) − EV(no investment, 2027)
26.  Cash out        Σ buyer payments 2026–2040, undiscounted
27.  ROI cash        ( Σ FCF(invested) − Σ FCF(no investment) ) ÷ cash out
28.  ROI EV          invest value ÷ cash out

Five inputs appear on the Inputs sheet, but no formula uses them. They are buffer b (0.05), convexity p (2), substitution premium (1), contribution margin (0.42) and exit multiple (12, itself labelled a sanity check). They are listed here so that a rebuild leaves them out.

Appendix BParameters

Every input, with its value and its label (§7.1). Values are at the default settings, with every scenario setting at Reference.

By crop

Yield index values

Only ratios enter the model, so the scale differs by case. Label: Sourced (CropSuite runs).

Shared by all crops

Set by the scenario settings, at Reference

The value of every setting at each level is in §6.2.

Appendix CScenarios and test configurations

Test configurations

A rebuild must reproduce the validation targets in Agri_Model_Data.json: company value doing nothing, value created on both routes, what you commit and both multiples, for all six crop and climate cases with every setting at Reference, and for every setting moved alone from Reference to Favourable and to Adverse. The workbook's own validation (§7.4) checks the same figures on each rebuild.

Price spike patterns

Each pattern is a multiplier on the world price trend for each year from 2026 to 2040. Scenario 1 is flat at 1.00 throughout. It is used only for testing and is not offered as a level of the price swings setting.

The peak price column includes each crop's own price trend, which comes from the climate. That is why cocoa's peak price is higher than its reference price times the multiplier alone.

The multipliers are set against today's price level, not a normal level from before the crisis. This is because the reference price is already tied to today's spot price. Setting them against the older level would count the rise already in today's price twice.

The record rule sets the height of each crop's patterns. At a normal price of 1.0 times, the two-spike pattern must match the crop's real price record. The shape of each pattern is kept. The part above 1.00× is scaled by one factor per crop, so years without a spike stay at exactly 1.00×.

Before scaling, coffee's patterns peak at 1.50×, which would put its two-spike price at USD 10,800/t, 11% above the record. Before scaling, barley's follow a recent swing of 21.5% and peak at 1.27×, well below the 2022 event. Scaling brings both under the same rule as cocoa.

Appendix DCost rates and sources

Rates are applied per hectare of programme area, which is set by the site (§2.4). A change of site changes programme cost in proportion and leaves the rates unchanged.

Appendix EHow results are shown

How the tool presents the figures in this document. None of it changes a result.

The tool's summary

The tool's summary shows both routes side by side, funding it yourself and buying credits, with a plus or minus sign on every figure.

At the defaults for Nawa, you commit USD 1.7bn and USD 1.4bn, and you create USD 8.0bn and USD 8.4bn.

Display rules

• A return is written as a number of times what you put in only when the commitment is at least USD 1m, the amount is not below zero, and the multiple is between 0.1 and 20. Otherwise the card shows the value created and what you commit as two separate amounts.
• An amount below USD 5,000 reads "none". An amount from USD 5,000 up to USD 50,000 reads "less than USD 0.1m".
• Scenario settings are shown as Favourable, Reference or Adverse. The tool's summary names the settings that differ from Reference, or reads "all Reference".
• The tool reads left to right through a chain of steps from the landscape to what the company is worth (§6.1). Each step shows the figure doing nothing and with the programme.