Every entry idea tested from 2026-09-17 to 2026-10-25, what it showed, and the resulting list of what NOT to trade. Real cost = about 0.2% per round trip (measured from broker fills 10/7). "Year 1 / year 2" = tested on two separate halves of the data. Updated 2026-10-25 (sections 7-9 added: situations, the EV3 chart setups, exits, gates, account simulations incl. 10-year out-of-sample). Scripts and full logs are named in each row. No look-ahead = decisions use finished bars only and act at the next price (two earlier results broke this rule and were corrected -- see section 8).
| # | Do not buy when... | Evidence | Strength |
|---|---|---|---|
| A1 | Price is more than ~2% above VWAP (stretched) | Stop hit first 40-48% vs 25% below VWAP, in 3 separate studies (9/26-9/27). Loosening the live 0..+2% band to +3% or 0.3x ATR: the extra trades lose -0.26% / -0.29% each over 2 years (10/9). | strong |
| A2 | The stock is calm: daily ATR under ~8% | Live v2 buys over 2 years: ATR < 8% = -0.16%/trade (1,482 trades, -$3,914) in every quarter; ATR >= 8% = about break-even (+0.02%) (10/10). Not enforced yet -- your decision. | strong |
| A3 | The volume rule fails (volume so far < 1.5x normal) | Looser volume 1.2x / 1.0x / off: extra trades lose -0.21% to -0.23% each, both halves (10/9). The 10/9 near misses (ABSI, CLF...) were winners in hindsight only. | strong |
| A4 | The rise is not steady (15-min smoothness < 0.7, straightness < 0.5) | Looser 0.6/0.4 or 0.5/0.35: extra trades -0.23% to -0.33% each (10/9). | strong |
| A5 | Price is already far above the open (more than 3%, unless a confirmed real breakout) | Loosening to 0.5x / 0.75x / 1x ATR: extra trades -0.18% / -0.27% / -0.26% (10/9). 20-day study: buys 3-10% above open -$624 (9/26). APPS bought at +5.6% reversed at once (9/25). | strong |
| A6 | Stock is falling on net selling (down >= 1% with sellers dominating) | Keeps falling: -0.58% / -0.70% over the next 60 min, both halves, 660 stock-days (flow study 9/28). Same in the VWAP-reclaim study (9/29). | strong |
| A7 | Late in the day (after ~14:30) | 2-year v2 replay: buys after 14:30 median -0.36%/trade; 14:30-15:00 negative both years (10/8). Late +2% touches almost never reach +5% (4%) (10/8). Late Vs (after 13:00) recover only 7-9% of the time (10/11). | strong |
| A8 | The first 15 minutes (before 9:45) | The 15-bar wait blocked SAIQ / ITG / MAX spikes that then collapsed (SAIQ -23%) (10/9). 20-day bar study: 9:30-10:00 entries -$441 (9/26). | medium |
| A9 | Heavy volume spikes (last-15-min volume > 5x normal, climax bars, acceleration at 10:00) | Heavy volume vs normal at 11-12 -> fade (AUC 0.42); high acceleration at 10:00 -> fade (AUC 0.41); climax events -21% to the close (5 cases) (speed study 9/28). Heavy-volume V recovery -8 to -10% (group study 9/29, small n). The X1 climax EXIT is built on this and passed 2 years. | medium |
| A10 | Gap over ~10% | Big-winner study: gaps over 10% faded about -8% from the open on average; gaps 5-10% were the best (9/27, 21 days). | medium (small) |
| A11 | A lower low by 10:30 | Only 2% of those became +5% winners vs 58 of 63 winners holding a higher low (big-winner study 9/27). | medium (small) |
| A12 | Spread over 1% or under 2,000 shares/min | Kept as live guards (spread since 9/28, liquidity since 10/7). Wide spreads eat the small gains (real cost already 0.2%). | rule kept |
| A13 | A V that is far below VWAP, late, or the 3rd+ V of the day | Opening-spike Vs: more than 5% under VWAP -> 80% stay weak, 7% recover; V after 13:10 -> 7-9% recover; 20th V -> 9% (2 years, both years equal, 10/11). | strong (as a filter) |
| A14 | Flat bars with trades right at the low (price pinned, O=H=L=C) | 40% keep falling vs 27% without flat bars (125k Vs, 10/11). Not a base -- usually sellers working the price. | medium |
| A15 | Breakout buys (vs pullbacks), and re-buys below a losing trade's entry | 20-day bar study: breakouts -$532 vs pullbacks about 0 (9/26). Re-entry only above the losing trade's entry was better on all 4 test days (9/23, live since). | medium |
| A16 | SEC filing: offering / dilution / delisting notice | Recording only since 10/7 (BETR delisting notice 10/9 while on the watch list). Not tested yet -- needs the 10-year tag backtest. | untested |
| Date | Idea | Result | Why it failed |
|---|---|---|---|
| 9/22 | Live entry-quality score (gate at 55) | Looked better on 2 days (in-sample), replaced later | 2 days, weights fitted on the same days |
| 9/23-9/28 | Setup J: trade plan, R:R >= 2, context rules, 5-day levels, slope rule | Mixed: +$145 on 6 days tick, -$98 / +$103 on 20 days bars; replaced by v2 on 9/29 | Results flipped between test sets; few trades |
| 9/26-9/27 | Shadow PASS gate (entry_rules: VAS, imbalance, relative strength, volume trend) | 20-day tick: J +$21 vs gate +$6 | Features flat over 64k snapshots; only VWAP distance mattered |
| 9/27 | trade1 entry_rules (620 picks, hold to close) | 311 up / 309 down, -$1,275 | No value separates rising from falling picks (AUC <= 0.56) |
| 9/27 | Buy early leaders at 10:30 | -$517 to -$1,610 over 21 days | Leaders by 10:30 are not more likely to keep going |
| 9/28 | Speed / smoothness signals (move >= 2x normal + smooth) | Held to close -0.6% to -2% vs time-matched base | 5-stock pattern did not generalize (654 stock-days) |
| 9/28 | Buy volume / net buying / "quiet demand" | AUC 0.49-0.50 for the next 30/60 min | Buy volume adds nothing beyond price |
| 9/29 | VWAP reclaim after an early spike (AGEN idea) | Every variant lost, both halves (-0.5% to -1.9% with stop) | Pullbacks below VWAP on selling keep falling |
| 9/29-9/30 | Playbook groups at 10:15 (ALLOW / AVOID) | In-sample good; 6-month out-of-sample ALLOW +0.00% vs AVOID -0.03% | Overfit to 23 days |
| 10/1-10/4 | Analog matching / Star Follow buys / reference trader / playbook advisor | All versions negative (-274% to -456% sums); advisor lowered P/L (+$118 -> +$69); Star buys -$197 / -$275 | Similar days predict MOVEMENT, not direction |
| 10/3 | Opening-range breakout (ORB, 12 variants, 503 days) | All lose (best -0.09%/trade) | Same as breakout finding |
| 10/7-10/8 | Pick only the best 1-5 trades per day (8 metrics) | At real cost: break-even at best (v3 score top 1/day: -0.08% / 0.00%) | Edge per trade smaller than cost |
| 10/7-10/10 | Deep-V entry (10 years) | Looked +0.16% to +0.22%; after fixing a 2-bar look-ahead: -0.07% to -0.10%, same as luck | The edge was future data |
| 10/8 | Buy the moment a stock touches +2% / +3% over the open (7 styles, 5 years) | Every style loses at 0.2%; the ">2% over VWAP + ATR >= 9.4% + early" pocket failed at checkpoint prices | Touch is common (38-49% of stock-days); no direction info |
| 10/9 | Group filters at 9:45-11:30 (16 groups) | Only "winner" group +0.22% train -> -0.17% test | Overfit |
| 10/9 | Group + up-trend confirmation (21 groups x 5 rules: new high, VWAP reclaim, V turn, 9:45 break, smooth trend) | 0 groups with an edge even in training; -0.15% to -0.45%/trade | After confirmation only ~1.3% move left |
| 10/9 | Looser v2 rules (12 variants: volume, steady rise, extension, VWAP band) | Every variant worse; extra trades -0.18% to -0.33% | The rules block losers (see list 1) |
| 10/10 | Buying pressure at the V turn (buy share 3/5 min, pressure shift, sell dry-up, net flow) | Reversal AUC 0.53-0.56, winning trade AUC 0.50; all filters -0.08% to -0.24% | Pressure seen too late; stop under the low too wide |
| 10/11 | Flat consolidation before the recovery, then buy > sell | -0.19% to -0.22%/trade, same as control; flat near the low = MORE falls | Flat bars = no trades / sellers pinning price |
| 10/11 | Style A (opening spike -> collapse -> V) recovery prediction | Recovery predictable (AUC 0.77) but trades lose (-0.09% to -0.14%); best filter overfit (+0.26% -> -0.09%) | The Vs that recover are shallow: too little room to pay the cost |
| 10/13 | EV3 (entry version 3) -- the user's VWAP + buy-flow entry: close >= +1% over VWAP (over +3% needs a confirmed break-up); path A gradual rise with buy volume last 3 min > 3 before; path B absorption (sells > buys 3 bars, price holding) then buyers; exit = 1/4 give-back (15c or 1%), volatility stop | 6 variants, 510 days: -0.21% to -0.24%/trade both years (gross ~0.00%), 43-55 signals/day; absorption no better (-0.21%); first signal per stock-day -0.22%; best slice 9:45-10:30 first signals -0.19% (gross +0.01%) | Conditions are common and carry no direction: same as random after the cost |
| 10/21 | C3: 3 up bars at the signal / wait for 3 up bars | No edge either way; only "buy share falling" looked useful | 3 up bars are common in winners and losers alike |
| 10/21 | Playbook buys (6 months, playbook as the buy signal) | No edge | Same as the 10/1-10/4 analog work |
| 10/21 | Unusual buy-share jump -> rise >= 1% | No edge | Jumps happen in fades as often as in runs |
| 10/21 | Dynamic situation map (per-minute situations A-H, 2 years) | No cell beats the baseline (+/-1% bracket) | Situations describe the moment, not the direction |
| 10/22 | Momentum / flow entries (volume + buy share + price all rising after 5 min) and "overrule" conditions for the situations | No edge on the 5 big losers and on 6 months | Same signals in winners and losers |
| 10/23 | Chart setups R1, R2, R5, R6, R8, R9 (practice charts; 2 years) | R2 / R6 = no better than a random minute (dropped); R1, R5, R8, R9 negative with every exit | See section 7 |
| 10/23 | E1 sell block (hold the give-back while buyers in control, JOBY/DERM) | First test +0.096%/trade, after fixing a look-ahead -0.065% (worse than no E1) | The hold checked the bar's own low (future data) |
| 10/23 | Dynamic give-back (+2% while buyers strong, +1% when they slow) and the "fade switch" GB2 -> GB1 | All worse than a fixed +2% arm (best +0.080% vs +0.108%), year 2 negative | Buy share dips right after every bounce buy, so the switch brings back the early sells |
| 10/23 | R1C first reversal after a flush (QUBT 5/20/25) | 284 buys, -0.20% to -0.35% with every exit, both years, every ATR group | Kept out of EV3 (the user: no changes to R1C) |
| 10/23 | X1 exit always (stall at the top / fade into resistance) and the climax sell | X1 +0.103% but y1 +0.512 / y2 -0.226; climax sell fired 0 times on the EV3 trades | X1 rides fat tails; climaxes come after GB2 has sold |
| 10/23 | N3 for R1A measured from the flush low (instead of 15 min) | Worse in year 2 (-0.221%) | Kept the live 15-min N3 |
| 10/24 | "Buyers in control for the last 5 min" (B55 / B60 / B4of5) or absorption as an EV3 gate | No difference or opposite in the two periods; absorption 7-16 buys in 2 years. Only R4 improves in both (B4of5: +0.46% / +0.75% kept) | Reversal setups buy while sellers are still there |
| Item | Evidence | Status |
|---|---|---|
| Scanner for EV3 (10/25): the OLD score (v1) top 30 | EV3 combination #1: +24.1% on old-score picks vs +10.3% on v3 picks; Star's v1 + 5-min rescans -4.8% (all setups) | trade bot from 10/26 |
| Scanner score v3 (daily ATR %, gap size, premarket volume) | Better than the old score in every one of 11 years (10 years, 399k stock-days) | trade bot; Star bot back on the old score from 10/12 (A/B) |
| The v2 entry rules as they are (volume, room, buy share, steady rise, VWAP 0..+2%, max 3% over open, 15-bar wait, spread, liquidity) | Every loosening tested worse (10/9) | live both bots |
| Volatile stocks only (ATR >= 8%) | About break-even vs -0.16%/trade on calm stocks (2 years) | not enabled -- decision pending |
| Limit orders at the ask (buy) / bid (sell) | Mid limits filled 0 of 8; ask limits 3 of 3 in ~5 s (10/8) | live both bots from 10/9 |
| Exits: min stop 2%, smooth exit, X1 climax exit | 2% stop +$369 vs 1% +$282 (24 days); X1 confirmed on 2 years | live (trade bot) |
| Star's sells | Recent 7 days: better on 28 of 49 trades; 2-year test: worse | Star bot live; watching |
| Situation rules as "not now" gates (S1-S9, record only) | S2 held 88 buys -0.238%; S3 (in window) 49 buys -0.275%; S6 3% total -252.2%; S8 5% room total -267.5% -> -149.4% | record only both bots (S3 window mode, S8 5%) |
| EV3 no-trade gate N2 (under the previous day's low) + N3 (15-min buy share < 45%) | 2 years: -0.052% -> +0.018%/trade, y1 +0.016 / y2 +0.020 (the 298 blocked -0.178%) | live trade bot from 10/26 |
| Give-back armed at +2% (GB2) instead of +1% | 2 years, no look-ahead, N2+N3 kept: +0.018% -> +0.108%/trade (y1 +0.212 / y2 +0.026) | live from 10/26 |
| C3 exit controller (GB2, X1 takes control when buyers made the run) | +0.108% -> +0.166%/trade (y1 +0.345 / y2 +0.022); TTGT +11.7% vs +2.9%, SVIA kept +3.4% | live from 10/26 |
| R1D spring (exempt from N2+N3) and R1E selling exhaustion | R1D (vol 3x) +0.328%/trade y1 +0.355 / y2 +0.312 (61); R1E (8 bars, +10 pts) +0.786% (16) | live from 10/26 -- but see section 9 (out of sample) |
| R3 + A3 (RVOL < 1.5x) + A20 (gap-down 0.5-3%) | The ONLY change that held on years never used to tune: R3 kept 2016-18 +0.41% (60) and 2024-26 +0.28% (135) vs removed -0.52% / -0.18% | LIVE 10/26 inside combination #1 (user 10/25): gated R1A R1D R4 R3, R1E off -- 2024-26 +24.1% / 2016-18 -3.9% |
The moves are real -- about 12 scanner stocks a day drop 3%+ and rise 3%+ from a V, and 33 of 96 rose 3%+ on 10/9 -- but every signal that can be known at the moment of buying (price shape, minute bars, trade-by-trade buy/sell volume, flat bars, groups, similar past days) separates winners from losers only after the fact. By the time a move is confirmed, too little of it is left to pay the ~0.2% cost and a stop. What reliably helps is avoiding (list 1), picking volatile stocks (v3 score), cheaper execution, and exits.
Still untested (needs new data being recorded now): order-book pressure (bid vs ask size, trade bot from 10/12), SEC filing tags (since 10/7), and more live days of Star's sells.
Live v2 buys, every signal, 503 days, 0.2% cost; each rule ON alone (real replay). Base: 2,012 trades, -0.133%/trade. No rule alone makes entries positive. Best: min_atr_pct 411 trades -0.007% (halves -0.20% / +0.18%); lower_low_by_1030 -0.109%; max_above_open / breakout_exception -0.129%; choppy_many_vs -0.120% (482 trades). Worse than base: heavy_volume -0.190%, recent_climax -0.186%, no_buy_after -0.152% (late buys were better than average here), max_gap -0.139%. No effect (v2 already covers): no_buy_before, max_vwap_pct, net_selling_drop. ALL ON: 18 trades -0.473%. On ATR >= 8% stocks: lower_low_by_1030 +0.087% (262), max_above_open +0.011% -> pair test pending.
Top-30 picks with news known at the open reach +5% far more often (39-41% vs 26-29%, both years; "why moving" 56-58%, guidance 50-52%, FDA 51-53%) but do not close better (offerings close about -2%). The bot's buys: any news -0.198% / +0.004% vs no news -0.189% / -0.158% (year 1 / year 2) -- not stable; price-target up/down buys lose both years. News = a MOVER signal (use in the scanner), not a buy signal.
20 random practice charts (reports/practice/practice_2210.pdf); the user marked the right entries, each became a setup module (trade/r*_setup.py). 2-year test with the give-back exit (no look-ahead, year 1 / year 2 per trade):
| Setup | From | 2-year result | Status |
|---|---|---|---|
| R1 retest reversal | WOLF 11/6/24 | -0.177% (94) | not used |
| R1A fast double bottom after a flush | HELP 7/3/25 | -0.083% alone; with N2+N3 +0.48%/trade (68) | EV3 |
| R1B undercut and reclaim (deep V) | PEPG 10/1/25 | 24 trades, mixed; account sim 8 trades -$2,905 | OFF from 10/24 (to rework) |
| R1C first reversal after a flush | QUBT 5/20/25 | -0.235% (284) | not in EV3 |
| R1D spring (buy the flush bar) | CERT 2/27/25 | +0.328% (61, volume 3x, no gate) | EV3 |
| R1E selling exhaustion at the low | RKT 2/28/25 | +0.786% (16) | EV3 |
| R2 calm base | SNAP / NVAX | = random | dropped |
| R3 opening push -> first pullback | UUUU 6/18 | ATR 6-10% only; +0.078% with GB2 | EV3 (ATR 6-10%) |
| R4 runner's second leg | TGEN 7/18 | +0.039% to +0.129% | EV3 |
| R5 / R8 / R9 (sell-off base break, pullback after reversal, climax reversal) | BBAI / RCAT / SOUN | negative with every exit | not used |
| R6 staircase | FRMI | = random | dropped |
| R7 pullback in a morning uptrend | CRML 10/3 | +0.039% | left out, to improve |
| Exit | Result (per trade, after 0.2%) | Note |
|---|---|---|
| Star's sells on the setup buys | -0.384% | worse than everything |
| 1% trailing stop / + noise2 | -0.152% to -0.187% / worse | |
| 1/4 give-back armed +1% (GB1) | looked +0.125% -> really -0.052% | look-ahead #1: the test sold at the level only if the bar later closed down |
| E1 hold (buyers in control) | looked +0.096% -> really -0.065% | look-ahead #2: the hold used the bar's own low |
| Give-back armed +2% (GB2) | +0.108% (N2+N3 kept) | live 10/26 |
| C3 = GB2 + X1 takes control when buyers made the run | +0.166% | live 10/26 |
| X1 alone / climax sell / dynamic or fade-switch give-back | see section 2 |
Account rules as live: 5 slots, 1% risk / 19% cap, re-entry above a losing entry, last buy 15:15, $100,000 start, 0.2% cost (sim_full.py, reports/ev3).
| Run | Result | Note |
|---|---|---|
| Original EV3 plan (no gate, GB1), 2024-26 | -9.4%, DD -13.2% | |
| Monday config (EV3 + R1D/R1E, N2+N3, C3), R1B on, 2024-26 | +13.1% (y1 +10.3% / y2 +2.5%), DD -4.4% | minus the 5 biggest trades: +$734 |
| Same, R1B off, quote flow / tick flow, 2024-10..2026-10-23 | +15.2% / +10.9% (second half +$1.6k / -$2.6k) | the flow method alone moves the result |
| Same config, 2016-11 .. 2018-11 (first 2 of the 10 years, never used to tune) | -13.6%, both years -$6.8k, 306 trades, win 38%, every setup negative | 58% of buys hit the stop (vs 43%); exits behave the same -> the ENTRIES are fitted to 2024-26 |
| EV3 on the Star bot's real lists, 10/5-10/23 (15 days) | +0.095%/trade on 15 trades vs Star's real -0.109% on 114 | small sample |
| Do-not-trade rules as EV3 gates, both periods | A20 gap-down blocks -0.38% both periods; A3 low RVOL -0.41% / -0.16%; R3 + A3 + A20 turns R3 positive in both | the AKS 4/12/2017 loser hit A2, A3, A8, A19, A20 -- EV3 checks none of them |
| Scanner score with EV3 (2024-26, tick flow): library picks / v3 top 30 / v1 top 50 / Star v1 + 5-min rescans | running -- v3 at 5/4/26: -3.2% vs the library picks +9.7% on the same date | pending |
Lesson of this round: every rule chosen on 2024-26 looks good on 2024-26. The 10-year data (flow being downloaded) is the referee: a change is only kept when it helps on years it was not built on.