Demand forecasting in an Iberian retail context
Hourly demand forecasting for an Iberian retail book is a different problem from the forecasting most teams inherit from generation planning. Weather, calendar, tariff shape and CUPS mix interact in ways that simple seasonal models miss, and the cost of being wrong on a single afternoon shows up in margin within 48 hours.
Demand forecasting in a retailer is a different problem from demand forecasting in a generator. A generator forecasts to dispatch. A retailer forecasts to price and to purchase. The decision horizon is different, the data sources are different, and the cost of error lands in a different P&L line.
This note sets out how the forecasting problem actually shapes up for an Iberian retailer, and what a working model needs to cover.
Why simple shape forecasts under-perform
Most retailers under 500 GWh inherit a forecasting approach that is some combination of seasonal averages, day-of-week shape and a manual override for known events. This works in a stable book in a stable price environment. It under-performs in two specific situations:
High-PV residential books on cloudy afternoons. A Madrid retailer with 30% of its book on PV-rooftop self-consumption tariffs has a demand shape that swings materially with cloud cover. A simple shape forecast does not capture this; the result is over-purchasing on cloudy afternoons and a margin loss on the day-ahead clearing.
Heat-sensitive SME books in a heatwave. A retailer with a 40% SME mix in Andalusia has a demand shape that spikes 25-30% above the seasonal norm on a 38°C day. A simple shape forecast under-purchases and the retailer takes a settlement hit on the REE balancing market.
The combined effect of a few of these days in a year is typically 200-400 bps of gross margin on a 500 GWh book. That is real money.
What a working model has to cover
Five inputs matter.
The hourly weather forecast at the customer-postcode granularity, not the city level. A book spread across Aragón needs different forecasts for Zaragoza, Huesca and Teruel; aggregating to "Aragón" loses information.
The customer-mix shape. Residential PVPC, free-market residential, SME on bilateral tariff, and corporate structured products each have a different demand shape. The forecast should be a weighted aggregate, not a single shape applied to the whole book.
The calendar. Public holidays in Spain and Portugal are regional. A national-holiday flag is necessary but not sufficient.
Tariff-shape effects. Customers on time-of-use tariffs (P1-P6 in the Spanish framework) have a different response to price signals than customers on flat tariffs. As the time-of-use share of the book grows, the forecast needs to model this explicitly.
Recent metered data. Most retailers receive metered data from the DSOs with a 30 to 60 day lag. The forecast should use that data once it arrives to recalibrate; in the interim, it relies on settlement estimates.
What you do not need
A few patterns over-promise relative to what they deliver:
- A deep-learning model for a 500 GWh book. A well-specified linear model with proper feature engineering gets within 1-2% of what the more complex approaches deliver, at materially lower operational cost.
- Real-time weather updates faster than 6-hourly. The day-ahead market clears at 12:00 CET for next-day delivery; a weather update at 11:00 is useful, an update at 11:55 changes nothing.
- A forecast that updates every five minutes. The market cadence does not support sub-hourly action; an hourly granularity is what the downstream purchasing workflow can use.
How this lands in operations
A working demand forecast feeds three downstream workflows:
- The day-ahead purchasing decision for OMIE. Hourly granularity, 24-36 hours ahead.
- The intraday purchasing decision for MIBEL intraday auctions and continuous intraday. Hourly granularity, 0-12 hours ahead.
- The hedging decision against MEFF futures. Daily or weekly granularity, 1-24 months ahead.
The model needs to deliver outputs in the right shape and the right cadence for each. A single forecast that tries to serve all three rarely serves any of them well.
Where to start
For most retailers under 1 TWh, the right first investment is a working hourly forecast for the day-ahead and intraday horizon. The hedging horizon can be served initially by a simpler quarterly model and upgraded later.
The purchasing pillar covers the forecasting deliverable as a 6 to 10 week engagement. Related: Operating against OMIE, Purchasing in MIBEL intraday markets.