Crypto daily baseline (BTC/ETH directional model)
hermes · confidence medium · status provisional
tags: crypto research baseline ml evaluation
Now
A provisional, research-only baseline authored by the AI-DLC Developer lane
(owner: originating Matrix request, no task ID supplied; no Kanban card accessed).
Artifacts live at /Users/dwayne/.hermes/artifacts/crypto-daily-baseline-20260910/
(baseline.py, test_baseline.py, test-results.txt, synthetic.csv,
synthetic-report.json, README.md).
Contract: at UTC midnight t, after the daily candle ending t is final, predict P(close[t+24h] > close[t]); ties are class 0. This is a close-to-close research target, not a claim that an order can fill at close[t]. Input is final, complete, consistently sourced daily BTC and ETH closes with timezone-aware UTC midnight candle END timestamps, at least 150 rows per symbol.
Model: separate fixed L2 logistic regressions per asset; past 1/3/7/14-day log returns plus trailing 14-return population volatility; train-only standardization; chronological 60/20/20 split with one excluded origin before each boundary; no hyperparameter search or calibration.
Verified with Python 3.14.4: five unit tests passed and the CLI produced JSON. Tests cover future-invariant features, target/purge boundaries, train-only fit/preprocessing, metric formulas, and end-to-end finite probability output. Input validation branches are implemented but not exhaustively unit tested.
Established: the software works on synthetic data. Not established: any predictive edge, calibrated probability, statistical significance, profitability, cost or slippage handling, drift monitoring, or trading readiness. No real market data was inspected, and real-data validation is blocked on a vetted point-in-time dataset. No orders, no data downloads, no dependency installs, no service changes.
Next interface (not dispatched): a data specialist supplies point-in-time final candles, source and quote-currency contract, close-time mapping, availability delay and revision policy; an evaluation specialist freezes the target contract, aligns dates and extends to walk-forward folds with block-bootstrap uncertainty.
Log
- 2026-09-13 (weekly synthesis): Created from the 2026-09-10 artifact README and
test results. Recorded as provisional research with the synthetic-versus-real
boundary stated explicitly, so a later reader does not mistake passing tests for
a market result. Source:
/Users/dwayne/.hermes/artifacts/crypto-daily-baseline-20260910/. , hermes.
Sources:
/Users/dwayne/.hermes/artifacts/crypto-daily-baseline-20260910/README.md · /Users/dwayne/.hermes/artifacts/crypto-daily-baseline-20260910/baseline.py · /Users/dwayne/.hermes/artifacts/crypto-daily-baseline-20260910/test-results.txt