- signal_booster.py: _compute_rates() now also computes a per-symbol win
rate (not just system-wide/direction aggregates); trade_executor.py's
Kelly sizing prefers it when the symbol has enough closed-trade history.
Added _as_datetime() to normalize closed_at across backends/drivers
that return either a real datetime or a string from raw SQL.
- trade_executor.py: volatility-filter and Kelly-sizing exception handlers
now log at warning level with the actual exception instead of silently
swallowing failures that affect how much money a trade risks.
- risk_manager.py: compute_partial_tp_levels() now returns all 3 levels
its docstring always promised (TP1 25% + TP2 35% + 40% trailing
remainder) instead of silently dropping the last 40%.
- trade_executor.py: compute_volatility_adjusted_size() was dead code;
now applied as a multiplier on the Kelly-derived trade_size (using
max_risk_pct=100 to reinterpret it as "scale the already-sized trade"
rather than "% of a bankroll", which would always collapse to this
pipeline's $5 floor at its actual dollar scale).
- walk_forward.py: grid-search fallback (when every combo is too sparse
to trust) now picks the combo with the most trades/highest PnL instead
of always the grid's arbitrary first entry. Raised MIN_TRADES_PER_FOLD
5 -> 15 for a more defensible statistical minimum.
219 backend tests pass (+10).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Le
2399d0cb0c