Projects 2026
Forecast Arena
A Monte Carlo baseline against two time-series foundation models, forecasting the next minute of market prices.
The game
Every minute, each model commits to a range for the next one-minute price before that price exists. When the price arrives, the round is scored. The three players are a Monte Carlo simulator (EWMA volatility and momentum, 10,000 paths drawn from the ticker’s own recent errors), Amazon’s Chronos-2 and Google’s TimesFM 2.5. The two foundation models are LoRA fine-tuned on minute prices.
How it’s judged
Beyond the game score, each model’s own 10th–90th percentile range is judged by pinball (quantile) loss, 80% coverage and direction accuracy. The test set is the last two trading sessions: 7,200 forecasts across 48 tickers in 15 markets, none of which the models saw in training. Input settings were chosen on a separate calibration session.
The repository also includes a portfolio Value-at-Risk engine that runs correlated Monte Carlo paths across 50 assets.
What it showed
Given raw prices, both foundation models collapsed towards “next price ≈ last price” and trailed the Monte Carlo baseline. Giving them one-minute returns instead closed most of that gap. The remaining differences between models are small, and direction accuracy stayed near a coin flip for all three. One-minute prices are mostly noise, so this is a careful comparison, not a trading strategy.