zpewNo results are published yet. The evaluation engine, the backtest and the cross-checks are finished, but the official runs have not been made. Results will be published here only from a clean commit, with that commit and the configuration stated next to every number.
What will be reported
- Each model's improvement over the RV21 baseline in MSE and QLIKE, against both the squared-return and Parkinson proxies, per asset and pooled.
- Results for both the sliding and the expanding window.
- VaR breach rates at 95% and 99%, with Kupiec test p-values.
- How often GARCH(1,1) fits end on the stationarity boundary.
- The GARCH(1,1) cross-check against the arch library.
Negative findings will be reported as plainly as positive ones. If a simpler model wins, that is the result.
Reproducing them
Every number will come from two commands, run on hashed data snapshots:
zpew-cli eval --config configs/default.toml
zpew-cli backtest --config configs/default.tomlEach run writes a run.json with the git commit, compiler, config hash and the hash of every snapshot used. See the README for the full steps.