Day-low support (buy CE) / day-high resistance (buy PE) retest — with approach-velocity "physics" filter

Research, not advice. Read-only study; no orders. Built 2026-10-16.

Question

Is there edge in buying ATM CE when Nifty retests the intraday day low (support), or ATM PE when it retests the day high (resistance)? Does approach velocity / deceleration into the level predict bounce vs break?

Verdicts

idea	verdict	why
Buy CE at day-low retest / PE at day-high retest (all 8 variants)	FAIL	Net-negative in train, tune and 2026 test for every variant. Best test (X10 Y20, rejection entry, 30m): −₹106/trade, n=208, vs placebo −₹210 (t=1.27, not significant)
Level "holds" more than random	FAIL (no level effect overall)	Overall hold rate is at or below random-level placebo (test 42–49% vs placebo 44–47%)
Physics: velocity/deceleration predicts hold vs break	PASS as a predictor, WATCH as a trading signal	Strong, stable in all 3 splits (slow approach: hold 61/51/57%; fast: 25/36/29%). But most of it is generic momentum: random levels show the same pattern (slow 49%, fast 30%). The day-low/high level adds ~+10pp only on slow/decelerating approaches
Slow + decelerating approach → buy the bounce	WATCH (prediction only), FAIL as an option buy	Hold 64% / 58% / 68% (train/tune/test) vs 50% random; but option P&L −₹296 / −₹258 / −₹49 per trade (30m)
Fast + accelerating approach → trade the break instead	FAIL	Breaks 80%+ of the time, but the break trade nets ≈₹0 train, −₹294 tune, −₹312 test
Fast rebound off the level (reject entry)	FAIL (2026-only)	Test +₹76/trade n=100, but train −₹293, tune −₹252 — not stable

Setup (no lookahead)

Data: Nifty 1m spot 2022-01-03 → 2026-10-06 (S3 copy of nifty50.parquet; read-only file not touched). Train 2022–24, tune 2025, test 2026 (Jan–Oct 6).
Level = running day low (CE) / day high (PE) as known so far; level must be ≥30 min old; price must have moved away ≥Y pts (20 or 40); retest = bar trades within X pts (5 or 10) of the level; entry window 09:45–14:45; first retest per side per day.
Entry: touch (close of touch bar) or reject (first bar within 5 bars closing back away from the level and with body away from it; no entry if it breaks 20 pts first).
Hold = price does not go >10 pts beyond the level within 30 min of touch. Break = it does.
Trade: long ATM option; SL = spot 20 pts beyond level; time exit 15/30/60m; EOD 15:15.
Premium: delta 0.5 + theta 0.02 pt/min (calibrated on real 1m ATM data: fitted delta 0.44 CE / 0.48 PE, ~0.01 pt/min drift). Real ATM 1m premiums exist only 2026-07-13→10-01; on those 149 trades the approximation correlates 0.95 with real fixed-strike P&L and is slightly conservative (approx −₹76 vs real −₹53/trade).
Costs (matched to prior trader-win-mine work): lot 65, slippage 1 pt/side premium, ₹50 round trip.
Placebo: same side, same distance to a fake level, random day in same split, time ±30 min, same exits; 10 draws/event. Velocity placebo: 15,000 random minutes with a fake level 5 pts ahead of recent move.

Main results (X=10, Y=20 most representative; full table main_vs_placebo.csv, grid.csv)

cfg	split	n	hold (Wilson 95%)	placebo hold	₹/trade 30m	placebo ₹	t vs placebo
X10 Y20 touch	train	914	45% (42–48)	52%	−266	−175	−3.0
X10 Y20 touch	tune	291	45% (39–51)	47%	−283	−171	−1.8
X10 Y20 touch	test	252	42% (36–48)	44%	−160	−188	+0.4
X10 Y20 reject	train	775	53% (49–56)	54%	−281	−191	−2.7
X10 Y20 reject	tune	245	54% (47–60)	51%	−316	−174	−2.1
X10 Y20 reject	test	208	49% (42–56)	47%	−106	−210	+1.3

Every variant (X5/10 × Y20/40 × touch/reject × 15/30/60m) is negative in every split. Costs+slip are ~2.8 premium pts/trade; the gross move after a retest is ~zero to slightly negative, so costs decide it.

Velocity / deceleration (the physics idea) — X10 Y20 touch, tercile cut-offs fixed on train

v5 = pts/min toward the level over last 5 bars (train terciles at 2.34 / 4.69); decel = velocity over bars −7..−2 minus last 2 bars (positive = slowing). | bucket | train hold | tune hold | test hold | random-level hold | |---|---|---|---|---| | slow approach | 61% (n305) | 51% (72) | 57% (69) | 49% | | fast approach | 25% (305) | 36% (124) | 29% (101) | 30% | | decelerating | 57% (305) | 56% (97) | 57% (90) | 47% | | accelerating | 32% (305) | 33% (120) | 28% (89) | 36% | | slow + decel | 64% (151) | 58% (38) | 68% (38), CI 53–81% | 50% | | fast + accel | 20% (187) | 27% (79) | 18% (56) | 32% |

Satya's intuition is right about direction: fast approach → break, decelerating → bounce, and it holds out of sample. Correlation with hold (train): v5 −0.27, decel +0.24, distance travelled −0.21, day range −0.19; time of day and level age ≈0. But the bounce after a "hold" is too small/slow to beat ATM option costs + theta + the 20-pt stop on losers.

Rule worth watching (not trading)

Slow-decel day-low/high retest: level ≥30 min old, price moved ≥20 pts away, retest within 10 pts between 09:45–14:45, 5-bar approach speed in the slowest third (≤ 2.34 pts/min) AND decelerating (decel ≥ -1.89). Expect ~60–68% holds vs ~50% random. As a long ATM option for 30m it still lost money in all three periods (2026: −₹49/trade, n=38). Possible uses: a filter to avoid fading fast moves into the day low/high, or a candidate for a cheaper instrument (futures, or selling premium) — not tested here.

Honest limits

Premium approximated except Jul–Oct 2026 (checked: corr 0.95). VIX not available in the data used — not included.
One event per side per day; 2026 test n≈200–250 per variant, subset n≈38.
Did not test futures, spreads or option selling, which may suit a 60%-hold signal better than ATM buying.
Test set looked at once for the pre-registered configs; the rebound and break-trade views are extra exploration and should be read as such.

Files: analyze.py, phase2.py, phase3.py, grid.csv, main_vs_placebo.csv, velocity.csv, placebo_velocity.csv, break_trade.csv, real_premium_check.csv, events_main.csv, events_reject.csv.