Reading a yield assessment: what P50 and P90 mean for a solar park direct investment

In the data room of a solar park project, next to the lease, the grid connection commitment and the building permit, sits a document many investors only skim: the yield assessment. Yet it decides more than any other paper, because the financial model takes from it the volume of electricity it will count on for 20 years. This article explains how such an assessment is produced, what the labels P50 and P90 mean, why the same P90 for a single year and for the whole term are two different numbers, and how to tell whether a provider's return forecast stands on the assessment or beside it.

What is a yield assessment, and why does it decide the return?

A yield assessment is the independent estimate of how many kilowatt-hours a specific photovoltaic plant at a specific site will feed into the grid per year, together with a statement of how certain that estimate is. It is prepared by an expert or a specialised engineering firm on the basis of the planning documents, that is, before the plant is built. The result is the specific yield in kilowatt-hours per kilowatt of installed capacity (kWh/kWp), usually for the first operating year and as a curve over the term.

Why this one number carries so much weight becomes clear on the revenue side: for a solar park with a fixed EEG tariff or a fixed offtake price, revenue is the product of volume and price. If the volume deviates permanently by one percent from the assumption, revenue deviates by one percent too, in each of the 20 years. Operating costs, lease, insurance and debt service are unaffected; the deviation lands entirely in the result. A yield assessment that overstates the yield by five percent therefore shifts the equity return of a debt-financed project by considerably more than five percent. What is left for the investor in the end hangs on this chain, as Investing in solar parks: revenues, costs and tax leverage at a glance shows in overview.

Where do the numbers in the assessment come from?

From a calculation chain with four links: irradiation at the site, irradiation in the module plane, losses in the plant and the resulting yield. It starts with global horizontal irradiation, which the assessor does not measure but takes from long-term datasets. In Germany these come from the German Weather Service (DWD), which combines ground measurements with satellite data into nationwide radiation maps, or from European satellite databases such as the European Commission's PVGIS system, whose default dataset covers the years 2005 to 2023 on a grid of roughly five kilometres. Which dataset the assessment uses and which period it covers is stated on the first pages and is the first check question.

In the second step, the simulation software converts horizontal irradiation to the tilted module surface; orientation, tilt, row spacing and horizon shading enter here. The third step is the loss chain: module temperature, soiling and snow cover, mismatch between modules, cabling, inverter, transformer and the internal power clipping where the DC/AC ratio exceeds one. The ratio between the energy actually fed in and what the modules would produce from the irradiation at nominal efficiency is called the performance ratio; according to Fraunhofer ISE's surveys, new, carefully planned plants achieve annual values between 80 and 90 %.

Each link in this chain is an estimate with its own uncertainty: the satellite dataset deviates from ground measurement, the long-term mean does not match any single year, the loss assumptions are empirical values, and the modules do not deliver exactly the nameplate rating from the datasheet. A good assessment quantifies these uncertainties individually and combines them into an overall uncertainty. That is exactly where the P values come from.

What do P50 and P90 mean?

P50 and P90 are exceedance probabilities. The P50 value is the yield that is reached or exceeded with 50 % probability, in other words the expected value of the forecast: in half of the cases the plant delivers more, in the other half less. The P90 value is the yield that is reached or exceeded with 90 % probability; only in one case out of ten does the plant fall short of it. The greater the overall uncertainty of the forecast, the further apart P50 and P90 lie. The gap between them is therefore itself a statement: it shows how sure the assessor is of the figures.

Mathematically, most assessments assume a normal distribution around the P50 value. The P90 then lies 1.28 standard deviations below the P50, the P75 0.67 and the P99 2.33 standard deviations below. Using our own simplified example with a P50 of 1,000 kWh/kWp and an overall uncertainty of 6 %:

Two misunderstandings can be cleared up with the table. First, P90 is not a guarantee: even the P90 value is undershot in one case out of ten, and an assessment is not liable for anything. Second, a low P90 value does not automatically mean a poor project; it can also mean that the assessor has honestly set the uncertainty high, for instance because the data basis is short or the site is complex. An assessment with a narrow gap between P50 and P90 is reassuring only if the small uncertainty is justified.

Why is P90 for one year different from P90 over 20 years?

Because part of the uncertainty averages out over the years and another part does not. The overall uncertainty of an assessment consists of two kinds. The first is weather variability: a single year can be considerably sunnier or duller than the long-term mean. How large this variability is can be seen in the German Weather Service's figures: in 2025, global irradiation in the German average was 9.4 % above the reference value for the years 1991 to 2020, the record year 2022 with 1,227 kWh/m² was higher still, while 2024 with 1,113 kWh/m² was only slightly above the mean. Over 20 operating years, however, good and poor years largely balance out; statistically, the weather variability of the mean shrinks with the square root of the number of years. The second kind is model and data uncertainty: if the satellite dataset is too optimistic at the site or the loss assumption too tight, that error acts the same way every year and never averages out.

In the example above, the 6 % overall uncertainty contains roughly 4.5 % weather variability and 4 % model uncertainty, added in quadrature. For a single year that gives the P90 value of around 923 kWh/kWp. Over 20 years the weather variability of the mean falls to around 1 %, the model uncertainty stays at 4 %, and the overall uncertainty drops to around 4.1 %; the P90 of the 20-year mean is then around 947 kWh/kWp. Good assessments state both values, often as "P90 (1 year)" and "P90 (10 years)" or "P90 (20 years)".

For reading the financial model this means: the one-year P90 is the stress value for the question of whether debt service can be met from current revenue even in a poor year; the multi-year P90 is the basis on which banks judge the viability of the financing over the term; and the P50 is the value with which the equity return can be quantified as an expectation, not a certainty. Anyone presented with a one-year P90 as a "conservative long-term forecast" is looking at a stricter number than necessary; anyone who conversely reads a 20-year P90 as a liquidity stress value underestimates the risk of a single poor year.

Which assumptions in the assessment should I check?

Above all those that do not come from weather statistics but are set. The P values are only as robust as the assumptions that precede the statistics, and these are documented in every assessment, usually in a table of input parameters. Six points deserve a look.

• Degradation: what annual loss of output does the assessment assume over the term? Fraunhofer ISE measured a mean degradation of nameplate output of about 0.15 % per year across 44 quality-assured plants in Germany; the large review studies by the US research institute NREL report a median around 0.5 % per year across thousands of datasets, with means above that. Assumptions between 0.25 and 0.5 % are common; Fraunhofer ISE itself uses 0.5 % in its cost models. An assumption below what the manufacturer's warranty permits as maximum loss needs an explanation.
• Availability: what share of the time is the plant assumed to be down, for maintenance, faults and grid work? Assumptions just below 100 % are common; an assumption of 100 % is not a forecast but a wish.
• Curtailment: does the assessment account for hours in which the plant may not feed in or it is not worth doing so? This concerns grid congestion as well as periods of negative exchange prices, in which new plants have no payment entitlement under §51 EEG. Some assessments state the yield only as available energy and leave the deductions to the financial model; then it must be checked there that they arrive. The mechanics are explained in Negative electricity prices: what they mean for solar and storage investors.
• Design: do row spacing, tilt, module type and DC/AC ratio in the assessment match the construction drawings? An assessment prepared for an earlier planning variant describes a different plant.
• Data basis: which radiation dataset underlies it, how many years does it cover, and how old is it? A short or old time series increases the uncertainty; an assessment that uses the mean of the 1990s tends to underestimate irradiation, one that uses only the sunny recent years overestimates it.
• Independence: who commissioned and paid for the assessment, and does the author have an interest in the sale? An assessment from the provider's own house is not an independent assessment, even if it is called one. A missing independent forecast is not a trifle but one of the warning signs that How to tell a trustworthy provider of energy direct investments summarises.

The degradation point deserves a second sentence, because it decides the later years: at 0.5 % per year the plant delivers around 9 % less in the twentieth year than in the first, at 0.25 % around 5 % less. Summed over the term, the difference is not small, and it acts in the years in which the loan is usually already repaid and the yield belongs to the equity. How modules age and what happens to the plant at the end of the 20 years is set out in A solar park after 20 years: continued operation, repowering or decommissioning?.

How does the assessment connect to the financial model?

Through exactly one line: the annual electricity volume. The first question for any return forecast is therefore which value from the assessment was carried over there, P50, P75 or P90, for one year or as a multi-year mean, and whether the degradation in it comes from the assessment or is the provider's own assumption. If the line is not shown, it can be recalculated: electricity volume in the first year divided by installed capacity gives the specific yield, and that must be found in the assessment. If it is higher, the provider is modelling a plant that, according to its own assessment, does not exist.

The same question arises for the bank. A financing through a promotional loan or a house bank loan is usually modelled on the P90, not the P50; the bank wants to know whether debt service is covered even in a poor scenario. A financial model that shows a high equity return on the P50 and just covers debt service on the P90 is not therefore unserious, but it answers two different questions, and both answers should be shown. Which documents a bank expects for this and in which order the financing is set up is described in Financing a direct investment: bank loan, KfW 270 and the pitfalls.

A third user of the assessment is often forgotten: the insurer. A loss-of-revenue or business interruption cover replaces lost revenue on the basis of an agreed indemnity basis, and that is derived from the forecast yield. Anyone insuring on an optimistic P50 pays premium for yield the plant will probably not deliver; anyone insuring on too low a value is underinsured in the event of a claim. The details of the indemnity basis are in Insuring a solar park or battery storage asset: which policies belong to a direct investment, and who holds them.

And for battery storage: what is a revenue assessment?

The counterpart with a different uncertainty. A battery does not generate electricity, it shifts it; its revenue depends not on the sun but on the price differences in the electricity market between charging and discharging hours, and on revenues from balancing services. A revenue assessment therefore estimates not kilowatt-hours but euros per kilowatt and year, and its input is not a weather dataset but a market model with assumptions about the generation fleet, the build-out of wind, solar and further storage, and the resulting price spreads over 10 to 20 years. Which revenue sources come together in this is explained in Direct marketing explained: day-ahead, intraday and balancing power.

The difference from a yield assessment is fundamental: the weather variability of a solar park is statistically well described, because there are decades of measurement data, and it averages out over the term. The price development in the electricity market is a scenario, not a random process with a known distribution; it can move in one direction over the term, for instance when many new batteries flatten the same price peaks. P values in the strict sense are therefore rare in revenue assessments; scenario bands with a base and a downside case are usual. In our own models we treat the assessment curve as the conservative case and run sensitivities alongside it, rather than adopting it as the expected value. What that means for return statements is in Battery storage returns: where the revenue comes from, and what is realistic; the risks behind it are set out in Risks in BESS direct investments, and how they are structurally addressed.

One thing they have in common: for storage too, a degradation assumption decides the later years, here that of usable capacity and efficiency over the cycles. Whether the revenue assessment models declining capacity or the nameplate capacity of the first year is among the first questions; the background is explained in The battery as a real asset: lifespan, degradation and warranties of a grid-scale storage system.

Checklist: ten questions for any yield assessment

1. Who prepared the assessment and who commissioned it? Is the author independent of the provider?
2. Which radiation dataset underlies it, which period does it cover, and was it calibrated against a nearby ground station?
3. Does the assessment describe the plant that will actually be built: module type, tilt, row spacing, inverter, DC/AC ratio?
4. What performance ratio results, and is it within the range comparable plants achieve?
5. How large is the stated overall uncertainty, and which individual contributions make it up?
6. Does the assessment state P50 and P90, both for one year and as a multi-year mean?
7. What degradation per year is assumed, and how does it relate to the module performance warranty?
8. Are availability, grid congestion and the hours without payment during negative prices accounted for, in the assessment or in the financial model?
9. Which value from the assessment is in the provider's financial model, and which in the model for the bank?
10. How old is the assessment, and has the planning changed since?

How we work with it

For every project we present to investors, an independent yield or revenue assessment is part of the data room, and our own financial model does not take the expected value from it but works with the P90 as the financing basis and shows the P50 alongside. We reconcile the assessment's input assumptions with the construction drawings, the degradation assumption with the manufacturer's warranty, and we deduct the hours without payment explicitly in the model. We give no return promises; we show which figure rests on which assumption. That is exactly what we review with you in a no-obligation initial consultation, preferably on the basis of a specific assessment.