Forecast the roof, not the region

On a mild Saturday in October, with a clear sky over Adelaide, the households Wirrandel supplies in South Australia draw less electricity from the grid at one in the afternoon than at three in the morning. More than four in ten of them have rooftop solar. Between ten and three their panels cover most of what they use and export the rest, and what remains for us to buy on their behalf is a shallow trough in the middle of the day.

For an energy retailer that trough is a hedging problem. We buy most of our customers' expected demand in advance through swaps and caps, so that a spike in spot prices cannot ruin us. If we expect more midday demand than arrives, we have paid a fixed price for energy we do not need and must settle the surplus against the spot price, which in a South Australian or Queensland spring is often below zero. The error costs us twice. My argument here is that a retailer should forecast its own customers' roofs from its own meter data, and that buying a regional estimate of solar output amounts to hedging somebody else's customers.

The midday that was not there

We learnt this expensively. In the spring of 2023 our residential forecast came from a respected specialist vendor. It modelled solar output for each region from satellite irradiance and the installed capacity recorded by the distribution networks, then scaled the result to our share of customers. The method was sound for a region. It was not sound for us.

Between September and November our midday load in South-East Queensland and South Australia came in 9 to 13 per cent below forecast on most mild days, and the quarter's loss against plan was A$41 million. When we looked for the cause we found three, all of them in our own book. Our solar and battery offers had attracted households with larger systems than the regional average. Solar connections among our customers were growing by about 1,100 a month, faster than the region's. And when we compared the networks' capacity records with what our customers' meters were exporting, 9 per cent of systems were recorded at the wrong size, some by a factor of two.

None of this was the vendor's fault. A regional model cannot know that a retailer's customers differ from their neighbours. The mistake was ours, in treating a forecast of the region as a forecast of our book.

Every solar home is a small weather station

The data to do better already sat in our platform. Every solar customer with a smart meter tells us, interval by interval, how much electricity left the house. On a clear day that export curve reveals the size of the system, the direction its panels face and how much the household uses while the sun is up. On a cloudy day the shape of the dips shows when cloud crossed the suburb. Across 330,000 solar homes our meter data amounts to a dense network of small weather stations, each measuring what matters to a retailer: power, not sunshine.

From late 2023 we rebuilt the forecast on that foundation. For each solar customer we estimate the system's effective capacity and orientation from past exports, correct the network register where the meter clearly disagrees, and combine those estimates with irradiance and cloud forecasts at the level of the postcode. Household consumption is forecast separately, so that the model can tell a sunny day from a family that has gone away for the weekend.

We ran the new forecast alongside the vendor's for two seasons before the trading desk relied on it. That was not popular. It meant another summer on the old forecast with a manual midday adjustment, which I could not fully defend. But a forecast that sets hundreds of millions of dollars of hedges deserves testing through a full cycle of weather, and our chief risk officer was right to insist on an independent validation every year.

By the spring of 2025 midday forecast error on our residential book had fallen from 11.4 to 4.1 per cent. The trading team estimates hedging losses about A$19 million a year lower than under the old approach. The capacity corrections have proved useful elsewhere: they now inform the planning of our virtual power plant, and we have passed the discrepancies to the networks whose records they concern.

Who should own the forecast

There is a reasonable view that forecasting is a commodity and that a retailer should buy it, as it buys weather data or billing software. For regional demand I largely agree. The market operator publishes forecasts of rooftop-solar output and operational demand, and several vendors do excellent work at that scale. Where I disagree is the last step: translating a region into a book. That step depends on knowledge only the retailer holds, about its own customers, and it is the step on which hedging losses turn.

Ownership matters for a less technical reason as well. When a forecast is bought, the people who carry its errors cannot see how it was made, and every bad week becomes an argument about whose fault it was. When it is built in-house, with the trading desk inside the design, the argument moves to the model itself. Our traders now challenge my team's assumptions every week, and the forecast is better for it.

I would add two cautions. The first concerns privacy. A model that can infer a household's panel size and daytime presence from its exports is sensitive, and its outputs should serve forecasting in aggregate, not marketing to individual homes without their agreement. We apply that rule and audit it. The second concerns humility. Several states now require rooftop systems to support emergency curtailment, through which exports can be reduced when grid demand falls dangerously low. Those events are rare and hard to predict, and our model still handles them poorly.

For directors the practical questions are short. Ask whether the company's demand forecast is built from its own customers or scaled down from a region. Ask what the forecast error was at midday and on the hottest evenings, not on average across the year. And ask who would notice if the installed-capacity records were wrong. In our case, until the spring of 2023, nobody did.