Published benchmarks

This directory is reserved for reproducible, out-of-sample benchmark outputs.

Do not commit a leaderboard produced from training data or a single hand-picked run.

The walk-forward harness can use a frozen Date,Close CSV with --input-csv. Each new run saves its input snapshot and SHA-256 in the summary. To reproduce a run, retain that CSV with the metrics and predictions. Live yfinance history may be revised later.

Arena input

The arena consumes an aligned CSV containing actual prices and one or more model predictions. Optional grouping columns can represent ticker and forecast horizon.

Example:

Date,Ticker,Horizon,OriginClose,Actual,LSTM,MultiTask
2026-01-05,SPY,3,99.7,100.0,100.2,100.1
...

Run

quant-forecast arena \
  --csv benchmark.csv \
  --actual-col Actual \
  --prediction-col LSTM \
  --prediction-col MultiTask \
  --group-col Ticker \
  --group-col Horizon \
  --time-col Date \
  --previous-actual-col OriginClose \
  --cost-bps 5 \
  --output reports/generated/arena.html

The generated table includes forecast error, directional accuracy, naïve-relative RMSE skill, and simple transaction-cost-aware directional strategy diagnostics.

OriginClose must be the price available when each forecast was made. For horizons above one session, shifting the target-date actual close would read a future price and distort every baseline-relative and directional statistic, so the arena rejects that input without an explicit origin close.

Annualised strategy diagnostics default to 252 periods per year; pass --periods-per-year 365 for daily continuous-market data.