Duckstradamus

Duckstradamus is a machine learning application that predicts New Zealand's wholesale electricity prices for every hour of the following day. Rather than displaying a price forecast alone, it recommends the cheapest time window to operate high-energy equipment, helping businesses reduce electricity costs. The project addresses the challenge of highly volatile wholesale electricity prices. Businesses with flexible energy usage, such as cold storage or manufacturing, can pay significantly different prices depending on when they consume electricity. Our goal was to turn complex forecasts into a simple, actionable recommendation. We trained the model using ten years of hourly data (2014–2024) from the New Zealand Electricity Authority. The dataset included wholesale prices, regional demand, generation by fuel type, lake storage levels, and weather data. After feature engineering, we created approximately 80 features. We developed both an XGBoost model and an LSTM model, then compared their performance. XGBoost achieved better validation results and was selected as the final model. For evaluation, we compared our model with a naïve persistence forecast, which predicts tomorrow's price using today's value. Although simple, it is a strong baseline for electricity price forecasting. Our model outperformed this baseline on 258 of 365 test days and also compared favorably with published research on New Zealand electricity price forecasting. The main challenge was predicting extreme price spikes, which accounted for most forecasting errors. Future improvements would focus on better spike prediction, adding prediction intervals to communicate confidence, and evaluating the model on larger electricity markets.