Models

Live feed disconnectedPAPER

AI Models

Train, validate and inspect the forecasting models. Every metric is an estimate on historical data — a model that beats its baseline once may still be wrong tomorrow.

LSTM

Recurrent net with long short-term memory over feature sequences.

GRU

Gated recurrent units — a lighter recurrent alternative to LSTM.

CNN

1-D convolutions over the time axis to detect local price patterns.

MLP

Feed-forward net over the latest feature snapshot and lagged returns.

Ensemble

Validation-accuracy-weighted blend of the four model forecasts.

Train models

Trains on historical candles with chronological splits — no lookahead. Runs execute in the background.

ADA-USDT · 1h · LSTM

Active runs

No active training runs. Completed and failed runs appear in the training history.

A trained model is a probabilistic forecast, not a promise. Only models that beat the baseline with statistical significance are promoted to champion.

Latest ensemble forecast

BTC-USD · 1h — probabilistic estimate, not a guarantee

Feature importance

Not implemented yet — and we will not fake it.

Per-feature attribution (gradient saliency for the CNN/MLP heads, permutation importance for all models) is a planned next improvement. Until it is computed from real trained weights, this card intentionally shows no importance scores — invented numbers would be worse than none.

Feature families the models currently see

returns & lagsSMA / EMA trendRSIMACDBollinger bandsATR volatilitystochasticOBV & volume z-scoreVWAP distancecandle shape

Which of these actually drive a given forecast is an open research question here — treat all model output as a statistical estimate.

Training history

Recent training runs with walk-forward validation metrics. Accuracy is measured on held-out data — treat every number as an estimate.