Scorecard and Paper Trading

Scorecard & Paper Trading

Scorecard (mosaic/scorecard/)

A SQLite-backed ledger of CIO recommendations and their realized performance.

Scoring algorithm (scorer.py)

For each pending recommendation row:

1. Resolve the day-0 close and the day-N close (N = 5 and 21 trading days, not calendar days).
2. forward_return_N = (close_N − close_d0) / close_d0.
3. alpha_5d = forward_return_5d − benchmark_return_5d (default benchmark 000300.SH, MOSAIC_BENCHMARK_TICKER override).

Price routing in _fetch_close:

• A-share index (e.g. 000300.SH) → pro.index_daily.
• A-share ETF (5xxxxx.SH / 1xxxxx.SZ) → pro.fund_daily.
• Otherwise stock/HK/US → pro.daily via _fetch_price_data.

This is what lets ETF recommendations be scored like stocks (so they show up in win-rate / skill).

Anti-lookahead

score_pending snaps "today" backward to the last completed trading day and only scores rows whose forward horizon has matured; immature/missing rows are skipped (missing rows are still marked scored so they leave the pending set).

Views / RPCs

• scorecard.append — ingest a daily-cycle state.
• scorecard.score_pending(cohort, today) — back-fill matured returns (idempotent; RPC, not a CLI subcommand).
• scorecard.list_skill — per-agent alpha / Sharpe / n.
• scorecard.win_rate — per-ticker directional hit rate: fraction where sign(action) · future-5d-return > 0, with sample size n.
• scorecard.latest_cio_actions — latest CIO portfolio.

Honest interpretation of "win rate"

Win-rate is the directional hit rate of this system's own CIO history (over already-scored rows). It needs several days of daily-cycle + back-fill to be meaningful (small n is unreliable). It is not a universal "this stock will go up" prediction.

Darwinian weights

darwinian.compute / darwinian.get_weights (mosaic/scorecard/weights.py) turn agent skill into evolutionary weights used downstream (e.g. by autonomous_execution).

Paper trading (mosaic/paper_trading/)

A bespoke paper-trading engine (the project dropped backtrader; this is not backtrader-based).

• Auth — register / login / logout / current_user (bcrypt-hashed; the trading extra). Session at ~/.mosaic/paper_session.json, DB at ~/.mosaic/paper_trading.db.
• Trading — buy / sell with T+1 settlement (bought shares aren't sellable same day), commission, and position tracking (get_positions, get_trades, get_account).
• Signal → order — suggest_order_from_signal turns an agent decision into a sized order (paper.suggest_order_from_signal).
• Cross-user auth is enforced on read + write paths.

CLI: paper register|login|logout|account|buy|sell|positions|trades|suggest.