Forecasting Financial Markets with Timestral
Summary
This article evaluates Mistral's Timestral time-series forecasting model (28M parameters) on financial data, focusing on stock returns, volume, and volatility. It finds that stock-volume forecasts show the strongest improvement, while returns offer no clear advantage over simple forecasts; fine-tuning provides little consistent benefit, and uncertainty intervals are often undercalibrated. The study emphasizes practical considerations like local CPU inference on a MacBook Air and cautions against overinterpreting model gains for investment decisions.