Model

Training performance and model diagnostics

Showing simulated data. Live Supabase integration coming soon.

Training Reward Curve

Episode reward over 150 training epochs (latest model)

In-Sample vs Out-of-Sample

Checking for overfitting -- smaller gap is better

MetricIn-SampleOut-of-SampleGapStatus
Win Rate40.336.110%
Avg PnL2.1-3.8281%
Sharpe0.520.1865%
Profit Factor1.020.8220%
Max DD-5.4-8.761%
Trades/Day3.43.19%

Max DD gap is 81% -- the model may be underestimating tail risk out-of-sample. Consider wider stoploss or drawdown-aware training.

Action Distribution

How often the model goes long, short, or stays neutral

Long38% · 119 trades

Avg PnL: -1.2 bps

Short29% · 90 trades

Avg PnL: -2.6 bps

Neutral33% · 103 trades

Avg PnL: 0 bps

Confidence Distribution

Model prediction confidence vs actual PnL

Higher confidence → higher PnL -- the model is well-calibrated

Feature Importance

Which inputs the model relies on most (top 10)

CVD 5m
18.0%+3%
Taker Buy Ratio
14.0%+1%
RSI (14)
11.0%-1%
OFI 5m
9.0%+2%
Volume Spike
8.0%---
Futures Premium
7.0%-1%
Trade Intensity
6.0%+1%
Large Trade Imbalance
5.0%---
Hurst Exponent
4.0%+1%
ATR (14)
3.0%-1%

Weekly Retraining History

Performance delta after each Friday retraining cycle

WeekWR BeforeWR AfterPnL BeforePnL AfterResult
W1136%38.7%-3.2 bps-1.4 bpsHealthy
W1238.7%39.1%-1.4 bps-1.1 bpsHealthy
W1339.1%40.3%-1.1 bps-1.8 bpsCaution