v0.2.4: the backtest metric was correct — and still easy to misread

[initial-d]

v0.2.4 is live: https://github.com/initial-d/ml-quant-trading/releases/tag/v0.2.4

An external contributor found that cost_drag was cumulative over the full backtest while neighboring return and volatility metrics were annualized. The calculation was correct, but the label made an invalid comparison look reasonable.

Thanks to @sergio12S and PR #47, the project now:

• emits cost_drag_cumulative as the primary field;
• preserves cost_drag as a compatibility alias;
• documents the unit and time basis of every validation metric;
• accepts archived reports in aggregation and audit tools;
• tests duration scaling, fee scaling, and legacy access.

Full technical story: The Backtest Metric Was Correct — and Still Easy to Misread

The useful question is broader than this repository: which backtest metrics have you seen misread because their unit or time basis was implicit? Examples from other engines, reports, or your own research are welcome.

[initial-d]

Follow-up: the reproducibility artifacts behind this release are now live on Hugging Face.

• Explore the 100,000-row deterministic synthetic OHLCV panel in the dataset viewer: https://huggingface.co/datasets/dddyym/ml-quant-trading-synthetic
• Download and load the 213-input MLP smoke-test checkpoint: https://huggingface.co/dddyym/ml-quant-trading-synthetic-mlp
• Paper page: https://huggingface.co/papers/2507.07107

Both repositories support anonymous downloads and are linked to the paper through arxiv:2507.07107 metadata. They contain no real securities or redistributed provider data, and they are meant for installation, inference, CI, and teaching smoke tests — not as alpha claims.

If you try either artifact, the most useful feedback is a reproducibility report: environment, exact command, and whether the viewer/load snippet worked. Successful and failed runs are both welcome.