Temporal Fusion Transformer detects anomalous market volatility with 0.82 F1-score
A developer built a market volatility sensor using a Temporal Fusion Transformer (TFT), shifting the goal from predicting stock prices to identifying when an asset is about to become abnormally unstable. Rather than forecasting a single return value, the model estimates the full distribution of returns across multiple quantiles (p10, p50, p90) for each asset and time horizon. An anomaly alert triggers whenever a real return falls outside the model's expected quantile interval, since volatility spikes often precede news rather than follow it. The TFT was chosen for its Variable Selection Networks, multi-head attention over time windows, and quantile regression output, achieving an F1-score of 0.82, an MCC of 0.80, and near-perfect quantile calibration. The key takeaway from the project is that reframing the problem toward estimating uncertainty — rather than chasing price precision — makes it both tractable and genuinely useful in financial contexts.
