Precision25— how it works

Stocks in the bottom 40% of their 52-week range are statistically prone to mean-revert upward — especially when forced selling has run its course. The model's job is to distinguish the genuine capitulation bottoms from the falling knives that keep falling.

Trained with a path-aware asymmetric label:

• Label = 1 if the stock hit +30% within 30 days AND did NOT drop -15% on day 1 after entry
• Label = 0 if it dropped -15% on day 1, OR never reached +30%

After day 1, the stock can wander all the way down -50% before bouncing — that's still a winner. The model learns the only thing it can reliably predict: whether a stock will whipsaw down in the immediate hours after entry.

Why these numbers? Step 62 swept (target, stop) tuples and asym horizon lengths on a clean universe (no leveraged ETFs); PA(+30%, -15%) with stop horizon=1d had the highest in-sample Sharpe (5.97) and lowest drawdown (-9.8%).

5-seed XGBoost ensemble (depth 4, 400 trees, decay-weighted training) trained on 78 features: technical (RSI, MACD, ATR, Bollinger, SMAs), positional (52-week range, distance-to-high), volume, options flow (deep-OTM calls, P/C ratio), macro (VIX, sector breadth, credit stress). Calibrated with isotonic regression on a 60-day holdout.

• Universe: 532-name stock universe (small/mid-cap focus)
• Filter: pos_52w_range ≤ 0.40 (bottom 40% of 52-week range)
• Top-2 picks per day by calibrated probability
• Flat position sizing — a step-83 audit found regime-based size multipliers add noise without improving risk-adjusted return
• Per-ticker reentry cap: 7-day cooldown
• Position size: risk-parity weighted across books — precision25 is the highest-volatility book, so it carries the smallest weight (~5% of the aggregate, ~$13 per pick)

No stops, no trailing. Each position is sold in 6 equal slices at fixed days from entry:

• Day 15 → sell 1/6
• Day 30 → sell 1/6
• Day 45 → sell 1/6
• Day 60 → sell 1/6
• Day 75 → sell 1/6
• Day 90 → sell remaining 1/6

Why partial exits? Spreading P&L recognition across multiple calendar dates dramatically reduces daily portfolio variance. A step-76 sweep of 9 hold/partial schedules found the 90-day 6-stage exit beat the original 60-day 4-stage on every walk-forward fold — in-sample Sharpe rose from 6.2 to 8.6, and the equity curve had fewer, shallower drawdowns. Same model, same picks, smoother curve.

The IS-vs-OOS gap is driven by Q4 2023 (a Magnificent-7-style narrow leadership rally where beaten-down stocks kept getting beaten while leaders ran away — that single fold ran roughly -10% CAGR / negative Sharpe). 6 of 7 quarters worked, 1 didn't. The ~-17% worst peak-to-trough is concentrated almost entirely in that one Oct-2023 → May-2024 window; in the 2024-2026 stretch the strategy had a single shallow -6% dip. Realistic deployment expectation: Sharpe ~6, CAGR 35-50%, calm -3% to -7% drawdowns in normal regimes, with rare -15%+ episodes if a narrow-leadership rally recurs.

Schedule-only upgrade (2026-05-15): the exit schedule changed from 60-day 4-stage to 90-day 6-stage; the model and label are unchanged. Research steps 76-89 also tested a tighter PA(+25%/-12%) label which lifts Sharpe further but with a deeper tail drawdown — that label change is deferred pending live validation of this schedule.

Pre-deployment audit: initial backtest universe contained 33 leveraged/inverse/VIX ETFs (MSTU, AMDL, SOXL, TQQQ, UVXY, VXX, etc.) which were inflating the apparent edge by riding leverage in bull regimes and adding catastrophic-decay risk on extended holds. All such ETFs are excluded from the production universe and the backtest above.

Vol Runup hunts asymmetric runups across 4 horizons (30/45/60/90d) with a winner-classifier re-rank, exits via 10% trailing stop or 90d max. Precision25 is more concentrated — single 30d horizon, single label, no second model, only 2 picks/day. They overlap on universe (both filter pos_52w_range ≤ 0.40) but diverge on entry timing and exit mechanics. Run together they should be partially correlated.

• A narrow-leadership rally can produce a -15%+ drawdown lasting several months (the Oct-2023 → May-2024 window did exactly this)
• Roughly a quarter of individual trades end deeply negative — this is normal for the strategy. A similar fraction of large winners (+50%+) carry the book
• No stops by design. You will watch positions go -30% intra-trade. Steps 78 re-tested every smart exit imaginable on the current label; pure HOLD with partial scale-out beat trailing stops, profit-locks and early targets every time.
• The median trade is only modestly positive — the edge lives in the right tail. Long stretches can feel flat before a cluster of winners matures.

Every weekday: 22:00 UTC (18:00 ET) the evening preview scan runs against currently-available features (typically through T-1 close), publishes top-2 picks to the Evening Preview panel. 10:00 UTC (06:00 ET) the final morning scan runs after Polygon's overnight options data has posted, replaces the preview with confirmed pending positions. 14:00 UTC (10:00 ET), after market opens,process_open_fills.py backfills entry prices at today's open and flips positions from pending to open. Saturday 14:30 UTC the model retrains on the latest week's data.