A Predictive Stock Entry & Exit SystemBuilt on OVTLYR behavioral analytics

Headline

The system identifies the current market regime (Crisis / Normal / Extended) from SPY breadth, scores each stock with a regime-aware composite, and exits each trade via a five-rule policy (T1/T2/T3/T3_EXT/CAP). Universe filters (avg volume ≥ 1,000,000 AND current price ≥ $5) keep illiquid micro-caps and reverse-split-to-death stocks out of the candidate list. Each pick is independent; position sizing is the trader's call.

Per-trade results — four validation slices

500 tickers were held out of all training and rule-selection. Combined with the time split at 2024-01-01, this produces four independent test slices. Slice C (unseen tickers × unseen time) is the strictest test — tickers the rules were never tuned on, evaluated over time the rules were never fitted to.

Generalization holds. Per-trade win rate stays in the 81.41–84.22% band across every slice. Slice C — 780 trades from 500 tickers the rules were never tuned on, evaluated on 2.4 years the rules were never fitted to — lands at 81.41% win rate / +6.43% mean / +8.59% median.

Exit distribution (Slice C — unseen-ticker holdout)

The four profit-take exits (T1/T2/T3/T3_EXT) fire only on positive-P&L conditions, so they have 100% WR by construction. The CAP timeout is the failure mode: ~20% of trades never get a profit signal and exit at the 120-day cap with an average loss of -22%. Tested mechanical stop-losses (-15% to -25%), trailing stops, breadth-divergence early exits, and ML-discovered multi-feature rules — every variant reduced per-trade Sharpe or portfolio Sharpe (often per-trade-positive rules cluster correlated losses at the portfolio level and double drawdown). The current exit set is empirically near-optimal.

Win rate by validation slice

Train (n=5,044)

84.22%

Slice A — unseen tickers (n=1,181)

84%

Slice B — train tickers × 2024+ (n=3,024)

82.67%

Slice C — both unseen (n=780)

81.41%

Random baseline (~53% drift)

53%

Per-trade first, portfolio second. The headline above and the slice table are per-trade metrics — what each independent trade looks like. Portfolio-level outcomes depend on how you deploy the rules; see Portfolio Deployment for bootstrap-simulated total return, Sharpe, and drawdown under two reasonable concurrency configurations.

What is OVTLYR?

OVTLYR (pronounced "outlier") is a behavioral analytics platform that quantifies investor sentiment for 3,300+ US-listed stocks. Its core output is the Fear & Greed oscillator — a 0–100 score per stock per day. Lower means more fear, higher means more greed.

The system on this page uses sentiment as the primary feature, with price-based signals (EMA deviations, sector breadth, order blocks) as confirmations.

Metrics are documented in the glossary.

Methodology

The pipeline

1. Feature engineering. Process 481 development tickers across all 11 GICS sectors. Extract Fear & Greed (F&G), EMA50 deviations, sector breadth, F&G rate of change.
2. Bottom identification. Find ~7,000 price bottoms using local-minima detection plus forward-return analysis.
3. Stock classification. Label each ticker Mean-Reversion (MR), Momentum (MOM), or Mixed (MIX) based on how it historically responds to fear signals.
4. Cross-ticker pattern mining. Find features that predict profitable entries regardless of ticker.
5. Model building. Iterate scoring models per regime through 254 versions.
6. Backtesting. 2024–2026 OOS, 2020–2023 IS. Skip-forward exit indexing after each trade prevents overlapping holds from being counted as separate trades.
7. Consolidation and CV. Cross-validate any candidate rule across year, regime, and classification splits before shipping.

Scale

• 3,300+ stocks in the OVTLYR universe. Universe filters: vol ≥ 1M + price ≥ $5.
• 500 tickers held out of all training and rule-selection.
• 11 GICS sectors represented across the universe.
• 4,985 out-of-sample (OOS) trades across the three OOS slices (Slice A + Slice B + Slice C); 5,044 additional in-sample training trades for reference.
• Training: Jan 2020 – Dec 2023. Time holdout: Jan 2024 – May 2026.
• Rule discovery: ~1,000 candidate rules tested with BH-FDR multiple-testing correction (q=0.10). Composite-scoring entries, classification, and the F&G<70 ceiling hold across every slice; tested exit alternatives all reduced per-trade Sharpe.

The Three Regimes

The market is split into three modes — Crisis, Normal, and Extended — based on how broad the market's uptrend is. Each mode has its own buy criteria. The mode is detected from SPY breadth, the percentage of S&P 500 stocks currently in uptrends.

SPY breadth — the share of S&P 500 stocks in an uptrend

Crisis

Normal

Extended

0%15%50%100%

Crisis0–15%

Deep fear. Low F&G, sentiment still falling, price well under the 50-day average, weak sector.

~30 days a year · highest conviction

Normal15–50%

Stock-level fear starting to turn. Below-neutral F&G ticking up, price under the 50-day average.

Most of the year · the everyday playbook

Extended50–100%

No new entries. Over half the index is already in an uptrend, so this is late in the move — the weakest forward returns of any regime. Existing positions are held.

16% of days · no entries taken

The strategy inverts across modes. In Crisis, panic-selling creates the best buying opportunities — everyone wants out, the stock is cheap, the recovery edge is largest. The mode is always checked first; everything downstream depends on it.

The n810 Trading Plan

The system has three pieces: universe filters (vol ≥ 1M, current price ≥ $5), the entry filter (F&G < 70 ceiling across all regimes), and a five-rule exit policy. Two rule refinements: trend-aware T1 suppression (skip T1 when the OVTLYR trend overlay is up) and MOM-aware T3_EXT suppression (let MOM-classified stocks compound past the first peak). See stock classification for the MR/MOM/MIX definition.

Universe filters (applied before any other rule)

The five exit rules

Each day after entry, these five rules are checked in order. The first one that matches sells the position. Four of them require the trade to be profitable; the fifth (CAP) is a hard 120-day deadline.

Each day a position is open, in this order

1. 1T1Sentiment back in greed, trade up 5%+ — unless the trend is still confirmed→ sell
2. 2T2Entered in deep fear, the bounce has faded, trade up 5%+→ sell
3. 3T3Opened in Crisis, sentiment recovered to 40, trade up 5%+→ sell
4. 4T3_EXTOpened outside Crisis, peaked at +10%, now fading — momentum names exempt→ sell
5. 5CAPNone of the above fired within 120 trading days→ sell

↺No match — hold, and check the same five tomorrow.

The rule set is validated across the four-slice protocol (train / unseen-ticker / unseen-time / both-unseen). Per-trade WR holds 81–84% across all four slices. Seven alternative exit policies were tested on the actual price paths (stop-losses from -10% to -30%, trailing stops from peak-10% to peak-15%, profit targets, F&G-greed exits). Every variant reduced per-trade Sharpe versus the T1/T2/T3/T3_EXT/CAP set. A prior version of the rules used per-sector F&G thresholds (Communication/Industrials at 60, Energy at 50, others at 40); a 2026-05-15 audit found those non-default thresholds were below the project's n=50 minimum sample size on the holdout, and uniform F&G=40 across all sectors improved win rate and Sharpe on every slice. The per-sector logic was retired.

Performance

By-year breakdown (Slice C: unseen tickers × 2024-2026)

The 2026 row reads as 100%, and it is not a result — it is an artifact of the window ending. A trade only enters this table once it has closed, and the exit rules close winners quickly while losers run to the 120-day cap. Positions opened in 2026 that are still open are therefore absent, and they are disproportionately the ones going badly. Marking those open positions to market pulls the 2026 win rate from 100% down to roughly 57%. Read the 2024 and 2025 rows, which have had time to resolve; treat any partial final year in this table as unfinished.

Why sector breadth carries more weight at its low tail

The entry score reads sector breadth — what fraction of the stock's own sector is in an uptrend. Until August 2026 that term paid a flat bonus anywhere at or below 20% and nothing more, which turned out to leave most of the signal on the table. Per-trade Sharpe keeps climbing all the way down:

It is a gradient rather than a cliff, and it is monotone on all four slices, so the boundary isn't a fitted artifact. It also isn't a proxy for something the score already knows: inside a single regime it roughly doubles Sharpe (Crisis holdout, breadth ≤5% versus 5–20%: 0.698 against 0.245), and it survives the same control against stock-level Fear & Greed. Per year, the bottom band beats the 5–20% band in ten of twelve comparable cells.

The score now grades the tail — an extra +1.5 below 2%, +1.0 below 5%, +0.5 below 10%. Across all four slices that raised per-trade Sharpe and surfaced more trades rather than fewer, which is the shape a real signal has: holdout Sharpe 0.287 → 0.302 on 705 → 780 trades.

Why the F&G<70 entry filter helps

Skipping new positions when F&G ≥ 70 is the single most consequential rule on the entry side. It blocks the "chasing greed" entries that produce most of the CAP-bucket loser tail. The threshold sits on the F&G=70 CAP-rate cliff: a sharp phase transition from 8-20% CAP-rate below 70 to 38-42% at or above 70.

Data Quality

The OVTLYR feed occasionally has bad-tick errors — a single-day price that's disconnected from the days around it and disagrees with other public sources. Without correction these produce fictional trades in the backtest. The pipeline scans every ticker at load time for spike-then-revert patterns, statistical outliers, and OHLC inconsistencies, and repairs high-confidence cases in-memory. Aggregate impact on the headline numbers is sub-0.1pp on win rate — the repair matters for correctness but barely moves the dial.

Portfolio Deployment

All numbers above are per-trade — what each independent trade looks like in isolation. A trader using The n810 Trading Plan still has to decide how to deploy it: how many concurrent positions, how much capital per slot. Those choices produce widely different portfolio outcomes from the same per-trade edge. The table below shows three operating points, run as 1,000-bootstrap simulations on the production trade list (within-date shuffle, 100% deployed, uniform sizing).

Every metric moves monotonically as slot count rises. More slots means lower headline return, higher risk-adjusted return, shallower max drawdown, and lower run-to-run variance. The 20 × 5% baseline returns +161% but with 44% std — one bootstrap might give +200%, another +118%. 50 × 2% gives +124% with 22% std, much steadier. Which point you pick is a risk-appetite question, not a correctness one.

Why deployment matters more than rule tweaks

We tested ~30 candidate rule changes — per-trade-Sharpe-positive exits, ML-discovered multi-feature stops, breadth-divergence triggers, regime-conditional entry filters. Most that improve per-trade Sharpe fail at the portfolio level: rules firing on market-wide signals (sector divergence, regime shifts) tend to fire on many positions simultaneously, clustering losses into single weeks and doubling drawdown. Per-trade Sharpe assumes trade independence, but a 20-position portfolio shares regimes, sectors, and calendars.

One lever survives portfolio validation cleanly: diversification. More slots mean structurally guaranteed correlation reduction, and the table above is monotone in it. The other candidate — ML-scaled sizing — did not survive re-testing, as above.

Caveats and known limitations

Tools

Four production scripts (analyze.py, scan.py, portfolio.py, daily.py) share the same entry/exit logic, universe filters, and rolling classifier. All four surface per-ticker historical performance under The n810 Trading Plan, so you can see how the plan has done on each name before you act on a new signal. The live tool and the backtest produce the same entry decision for any given (ticker, date).

The daily list is a standing pool, not a day's worth of new ideas

A name that satisfies the entry condition keeps satisfying it, and is republished every session it does. So the headline pick count is a stock, not a flow: on 2026-08-12 the list showed 29 picks, of which 2 had not been picks the previous session. Over 2026-07-16 to 08-12, 549 pick-slots came from roughly 18 newly-qualifying names per session. Each pick now carries whether it is new and how many consecutive sessions it has qualified.

That raises a fair question — is a name that has been sitting on the list for three weeks a staler buy than a fresh one? Tested by delaying entry until the condition had held for N consecutive days:

Standing picks are better per trade, not worse — monotonically, on win rate, mean, Sharpe and CAP rate, replicated on Slice B. Persistent qualification selects for sustained fear, which is the setup the plan wants.

Two things stop that from being a strategy. Waiting forgoes 87% of the trades (holdout total return 5,018 → 897), and the comparison is not like for like: surviving to day 20 means the condition never broke, and it breaks on recovery, so late entry quietly excludes the names that recovered fast. Read it as the persistent subset is good, not waiting is free. The practical takeaway is narrow and useful: a pick that has been on the list a while is not damaged goods.

---

Keep reading