Price forecasting for a renewable energy trading platform
Lumion, a New York energy platform, needed price forecasts accurate enough for its users to decide when to buy, sell and store power.
The problem
Electricity prices in deregulated markets swing with weather, demand, grid capacity, fuel costs and regulation. Users needed reliable 24-to-72-hour forecasts to time selling, buying and storage. Every approach the team had tried produced error rates too high to act on.
What we built
A deep-learning time-series model predicting hourly prices across NYISO, PJM and ERCOT, combining temporal convolutions with attention and ingesting historical prices, live weather, load forecasts, generation mix and gas futures. Daily automated retraining, a backtesting framework, and probabilistic output with confidence intervals for risk-adjusted recommendations.
The result
Mean absolute percentage error of 8.2 percent on 24-hour forecasts — three times better than before. Users saw a twenty-two percent average lift in trading profitability within a quarter. The capability helped secure a $4.5M Series A; coverage has since grown from three markets to seven.
