It's 15:00. The intra-day price forecast was right.
The trade still lost money.
A block of inverters tripped at midday. SCADA showed it. Nobody on the desk saw it until the volume came up short. The book was short, and the only way to close it was to buy in a market that was already climbing.
The price forecast didn't fail. Everything underneath it did.
Most trading conversations at renewable IPPs start and end with price. Where will day-ahead clear? Where will intra-day go? Is the spread worth taking? Those are good questions. They're just not the risk.
Five layers, not one
The price forecast is the top layer. Under it sit four more, and each one can turn a good trade into a loss on its own.
Asset. Availability, outages, derating, maintenance that ran over.
Weather. The forecast error, not the forecast. A cloud front two hours early.
SCADA. What the plant is doing right now, if anyone is looking.
Curtailment. The grid says no. The plant is fine, the sun is out, and the volume still isn't yours to sell.
Price. Day-ahead, intra-day, imbalance.
Orange: the layer most desks model. Teal: the layers underneath. All five feed the same scenarios.
Curtailment is the one people forget. You sold the volume day-ahead. Then an order comes in, and the plant can't deliver it. From the book's point of view, that's the same as a broken inverter. You're short, and you pay to close it.
The danger is the combination
Low output on its own is a nuisance. A high intra-day price on its own is an opportunity.
Low output, a high intra-day price and a short position, all in the same hour, is the loss.
And those don't arrive independently. When renewable output across a region drops below forecast, everyone is short at once, and intra-day prices move with it. Published work on German markets finds exactly this: wind and solar forecast errors push intra-day prices, and not in a straight line.
Each dot is one scenario. "Correlated" dots come from a model that keeps output and price moving together. "Independent" dots come from a model that treats them separately: they cluster in the middle and miss the top-left corner.
A model that treats each input separately puts too few scenarios in that top-left corner. It isn't wrong about any single input. It's wrong about how they move together. So it under-prices exactly the hours that hurt.
How the risk is actually calculated
The method isn't new. Banks and the bigger trading houses have run versions of it for years. What's different for a renewable portfolio is how much of the input is physical.
Step by step:
Align. Plant, weather, market, SCADA and curtailment data, joined on the same delivery interval.
Forecast ranges, not points. Output at P10, P50 and P90. Same for intra-day price.
Measure the errors. Compare past forecasts with what actually happened.
Keep the relationships. Low output and high prices travel together. The scenarios have to show that.
Simulate. Thousands of joint scenarios, plus named stress cases: outage at peak, early cloud front, curtailment order, price spike, thin liquidity.
Value the whole book in every scenario.
Backtest. Compare the risk you predicted with what you got. Recalibrate.
For each delivery interval, the P&L in one scenario is plain arithmetic:
P&L = DA price x DA position
+ ID price x ID position
+ imbalance price x residual volume
- trading and execution costsRun that across every interval and every scenario, and you don't get a number. You get a distribution.
Reading the tail
Red: the worst 5% of scenarios. VaR is where red starts. CVaR is the average of the red bars.
Three numbers come out of it.
Expected P&L. The average outcome. It tells you whether the trade is attractive. It tells you nothing about how bad it gets.
VaR. At 95%, the point where the worst 5% of outcomes begins. In the example above, that's about EUR 35k. It's the edge of the tail, not the maximum loss. People misread it that way all the time.
CVaR, also called Expected Shortfall. The average loss inside that worst 5%. Here, about EUR 45k. This is the number that matters in power, because price spikes and correlated plant failures make the tail much fatter than normal volatility suggests. Banking regulators came to the same conclusion and moved market-risk capital from VaR to Expected Shortfall.
Then stress loss: the book revalued under the named bad days, one at a time.
Don't add up the VaRs
Ten open positions. Ten VaRs. The tempting move is to add them.
Don't. VaR doesn't add up. Positions can offset each other, so the sum overstates the risk. Or they all share the same weather, so the sum can understate it.
Grey: each position's own VaR. Orange: the three added up. Teal: the book's real VaR, once when positions offset and once when they share the same weather. The orange bar matches neither.
The fix is simple. Add the P&L scenario by scenario across the whole book. Then read the tail of that.
It also changes how you judge a single trade. Every proposed trade gets three views:
Standalone. How risky is this trade on its own?
Book. How risky is everything we already hold?
Incremental. Book CVaR after the trade, minus book CVaR before.
A trade can look risky on its own and still make the book safer because it offsets something you already hold.
How big is too big
Once you have a risk budget, sizing stops being a gut call. The right size is the biggest position whose CVaR stays inside the budget.
Red: CVaR, the average loss in the worst 5%. Teal: expected P&L. Grey: the EUR 50k CVaR limit.
Expected P&L (teal) keeps rising with size. CVaR (red) rises faster. At 20 MW, the trade fits under the EUR 50k limit. At 30 MW, it doesn't, even though it still looks more profitable.
The model can't set that limit. The business has to decide: the horizon, the confidence level, the budget, and who approves exceptions.
The maths is the easy part
These methods are standard, and the formulas are brief.
The main challenge is data integration. Plant availability, weather, SCADA, curtailment orders, and market prices are often stored in separate systems, formats, or time zones, making alignment difficult.
If these data sources are not aligned to the same delivery interval, scenario analysis becomes unreliable, and VaR results are misleading.
This is a clear example of the Decision Lag. The intra-day window is limited to a few hours, and while the necessary data exists within the business, if it cannot be consolidated quickly, trades are executed based solely on the price forecast.
Therefore, the sequence is important. Establish a unified, governed view of plant, weather, SCADA, curtailment, and market data before applying the risk model.
The price forecast should initiate the risk discussion, but it must not be the sole consideration.
How does your desk currently assess position risk: solely by price forecast, or by incorporating plant factors as well?