The Independence Assumption Hiding in Your P90

The riskiest figure in a renewable energy financing is the one labelled conservative.

That figure is the P90, the annual energy yield a project is expected to exceed in nine years out of ten. Lenders size debt to it. Sponsors defend it in committee. It is the closest thing project finance has to a floor under a deal. It also rests on an assumption that the climate has stopped honouring, and a strengthening El Niño is about to make that assumption expensive.

In June 2026, the United States National Oceanic and Atmospheric Administration moved its alert status to an El Niño Advisory, confirming the event had formed and assigning a 63 percent chance of a very strong event by the 2026-27 northern winter.[1] For an Asia-Pacific portfolio weighted toward solar, wind, and hydro, that is not a weather headline. It is a signal about the shape of revenue over the next two years, and about whether the downside cases in your existing models were ever built to absorb it.

The assumption no one states out loud

A P90 is built from the statistical spread of historical weather. Take the expected output, the P50, apply an uncertainty band drawn from how much generation has varied from one year to the next, and read off the downside percentile that the debt schedule has to clear. The arithmetic is sound. The problem sits in what the arithmetic quietly assumes: that each year is an independent draw, and that a poor year carries no information about the year that follows it.

For most of the industry's history, that simplification was harmless. Weather was treated as noise around a stable mean, and noise averages out over a twenty-year tenor. The difficulty now is that the single largest driver of year-to-year variability in Asian renewable resource is not noise. It is a climate cycle that runs in multi-year blocks, and a model that cannot see the block cannot price the risk it carries.

ENSO does not deal bad years one at a time

The El Niño-Southern Oscillation, whose warm and cool phases are El Niño and La Niña, does not hand out one weak year and move on. It hands out runs. The most recent cool phase, which ran from 2020 to 2023, was the first "triple-dip" La Niña of the century, three consecutive years of the same regime and only the third such event since 1950.[2]

When the dominant driver of resource variability is itself serially correlated, a downside case built on independent draws understates the real exposure. The threat to a project is rarely a single disappointing year, which a debt service reserve account is designed to absorb. The threat is two or three poor years in sequence, which it is not. The reserve covers the first shortfall, thins on the second, and is gone by the third, precisely when the covenant test bites.

This is visible in the operating record, not only in theory. A study of onshore European wind by Hendrickson Renewables found a median P50 bias of negative 8.9 percent, meaning realised generation tended to land well below the central case used to raise the debt. Only 14.3 percent of the projects examined matched or beat their pre-construction energy yield assessment, and the bias showed little improvement in more recent vintages.[3] That is not a rounding error inside an uncertainty band. It is a structural tilt in the wrong direction.

India shows the same pattern with a monsoon signature. Analysis by the Council on Energy, Environment and Water found that national wind generation in the peak June to September season of 2020 came in 24 percent below the same months of 2019, with the western region down 29 percent and the southern region down 17 percent.[4] The western grid had to lift hydropower output by 12 percent to cover the gap.[4] None of this appears in a typical meteorological year file, because that file is constructed precisely to smooth such runs away.

The year-three problem

Picture an investment paper that clears committee on its merits. A twenty-year offtake, an investment-grade counterparty, a P90 with comfortable headroom over the debt schedule. Every box is ticked.

Year one comes in light. Year two, lighter. By year three the project is drawing down its reserve account and the lenders have asked for a meeting.

No one mis-modelled the turbine or the panel. The technical work was sound. The weather assumption was the problem, and specifically the idea buried inside it that bad years arrive scattered and independent. A reserve sized for one off year does not survive three in a row. That is the route from a fully banked, investment-grade deal to a covenant conversation, with no modelling error anywhere in the file, only a modelling assumption that the climate declined to respect.

The financial consequence is not abstract. A clustered shortfall pushes the debt service coverage ratio toward its covenant floor, forces a reserve top-up that competes with distributions, and, in the harder cases, triggers a cash sweep that holds equity returns hostage until generation recovers. For development finance co-investors who have written climate resilience into their own diligence standards, it is also exactly the failure mode their stress testing is meant to catch, which makes its absence from a sponsor's base case harder to defend each year.

Why Asia-Pacific compounds the problem

The point sharpens in this region because the exposures are correlated rather than diversified. Across much of monsoonal Asia, wind output, hydro inflows, and cooling demand all track the same seasonal circulation. A weak monsoon can suppress wind generation, draw down reservoirs, and lift peak demand in the same year, across the same portfolio. The 2020 Indian episode was a preview of the mechanism: when wind fell, hydro was called on to fill the gap, concentrating the stress rather than spreading it.

A book that looks diversified by technology on a slide can behave like a single bet once ENSO becomes the common driver. That is the precise risk that a year-by-year independent model is structurally unable to show, and it is the reason a regional portfolio needs its downside cases to share a weather chronology rather than each carry its own smoothed average.

The fix is methodological, not technological

The correction does not require new hardware or a forecasting breakthrough. It requires modelling weather the way it actually arrives.

In place of an averaged typical year, leading technical advisers now build synthetic chronological weather series that preserve how conditions sequence themselves, then stress the model against a reconstructed ENSO block: a three-year wind drought, a monsoon-failure hydro year, a clustered downside rather than a smoothed one. The tools exist and are in use. What is often missing is the discipline to insist on them in the committee paper.

Three questions separate a downside case you can bank from one that only looks conservative.

First, does the model assume bad years are independent? Almost all do, because the standard P90 method is built that way. If the downside case cannot represent a cluster, it cannot price the risk that actually breaches cover.

Second, has it been stress-tested against a real ENSO sequence rather than an average? A typical year is an average of decades, engineered to remove the multi-year runs that matter most. The useful test runs the asset through a bad sequence drawn from the historical record, not around it.

Third, is the reserve sized for one bad year or for several? This is where the first two questions land on the balance sheet. Most reserve accounts are built for a single off year. A clustered ENSO block drains them in order.

The number that protects a deal is not the central case in the headline. It is the downside the project can survive in sequence. With El Niño now formally underway and better-than-even odds of a strong event into 2027,[1] the next renewable paper that crosses your desk with a clean P90 and no clustered-downside scenario deserves one question before any other: what happens in year three, if year two was already bad?