Mega-Trading

Backtest Dashboard

Self-contained report generated from local training, MLflow, dataset split, and backtest artifacts.

Run
rtx
Mixture
rtx
Checkpoint
present
Steps
2,500
Train loss
3.3414
Validation loss
2.3500
Backtest loss
2.3379
Backtest ppl
10.3592

Artifacts

Checkpoint.mega-trading/data/runs/rtx/checkpoint.ptpresent
Metrics.mega-trading/data/runs/rtx/metrics.jsonpresent
MLflow DB.mega-trading/data/runs/mlflow/mlflow.dbpresent
Backtest JSON.mega-trading/data/evals/rtx/backtest.jsonpresent

Prepared Split

train54,695,345validation1,116,216backtest6,201,267
Chronological per-ticker split from prepared dataset metadata.

Training Curves

2.333 - 6.553
train loss | validation loss
Green: train loss. Blue: validation loss when evaluated.

Ticker Candles + Forecast

Green/red candles show sampled real Binance trades from a held-out archive. Blue line is the model-implied price path from generated relative-price tokens seeded on the same ticker stream.

BTCUSDT 2026-04

67938.1125 - 68284.4900
Green/red: real OHLC candles. Blue: generated model-implied price path.

ETHUSDT 2026-04

2100.0738 - 2106.8800
Green/red: real OHLC candles. Blue: generated model-implied price path.

BNBUSDT 2026-04

616.0000 - 619.5000
Green/red: real OHLC candles. Blue: generated model-implied price path.

Backtest Metrics

Loss2.3379
Perplexity10.3592
Top-1 accuracy49.23%
Top-5 accuracy69.42%
Scored tokens524,288
Distribution L10.1420

Stylized Facts

MetricRealGenerated
abs_return_autocorrelation_lag1-0.0001-0.0005
count16384.000016384.0000
excess_kurtosis3.094814.2966
extreme_fraction0.00000.0000
mean_bps-0.7012-0.8331
return_autocorrelation_lag1-0.0696-0.0057
std_bps0.62720.4405

Run Configuration

Vocab size579
Block size512
Sequence count62,012,828
Training backendaccelerate
Distributed strategyddp
World size1
Attention backendtriton
Compile enabledTrue