AI Research NVDA

NVDA 3-year up-streaks (≥5 days): forward 3-day returns vs unconditional baseline

A five-plus session green streak in NVDA’s recent tape does not leave the stock measurably “due” for a three-day pullback. Over the trailing ~3 years there were 21 such streaks; the mean forward 3-day return after those events was +0.75% versus +0.74% for the 749-day unconditional baseline, and a Welch t-test gives t≈0.012, two-sided p≈0.990. In short, the tiny observed gap is indistinguishable from noise.

The study resampled minute bars to daily closes, flagged days that ended a ≥5-day up-streak, and compared the distribution of close[t+3]/close[t]−1 to the unconditional universe. Sub-buckets swing wildly on very small Ns, so read the detailed charts and tests below — the full evidence supports the thesis that the “overdue for a dip” reflex looks like gambler’s fallacy, not a tradeable 3-day mean-reversion edge.

The research question

For NVDA over the past ~3 years, does a long green streak actually leave it 'due for a pullback' — after the stock closes higher five or more sessions in a row, are forward 3-day returns any weaker than the everyday baseline, or does the run just keep going? Thesis: forward returns after long up-streaks come out statistically indistinguishable from baseline, so the 'overdue for a dip' reflex is the gambler's fallacy rather than a tradeable mean-reversion edge.

How this was measured

Resampled NVDA minute bars to daily closes, restricted to the trailing ~3 years based on available data. Computed daily up/down and the length of the consecutive up-streak ending on each session. Defined an event when the stock closed higher for 5 or more sessions in a row; measured the forward 3-trading-day return as close[t+3]/close[t] − 1. Compared the event-day forward-return distribution to the unconditional baseline of all valid 3-day forward returns via a Welch t-test (unequal variance). Also summarized by streak length buckets (5, 6, 7+).

The key numbers

Window trading days
752
2023-06-30 to 2026-06-30
Up-streak events (≥5 days)
21
Anchors with valid t+3 close
Baseline anchors (all days)
749
All days with valid t+3 close
Event mean fwd 3d return
0.7469%
N=21
Baseline mean fwd 3d return
0.7380%
N=749
Edge (event − baseline)
0.0089%
Event median fwd 3d
1.0301%
Baseline median fwd 3d
0.6246%
Event fraction positive
57.14%
Baseline fraction positive
55.41%
Welch t-statistic
0.012
Positive favors event mean
Welch p-value (two-sided)
0.9904
p=0.9904 ≥ 0.05 → no statistically-clear difference

Reading the numbers

After 5+ consecutive up closes, NVDA’s mean forward 3-day return is 0.007469 vs a baseline mean of 0.007380 — effectively the same. There are 21 event anchors versus 749 baseline days and a two-sided p-value of 0.9904, so no statistically-clear difference.

The charts

NVDA forward 3-day returns after ≥5-day up-streaks
What this chart says

This histogram shows the 21 forward 3-day returns after 5+ day up-streaks, with a mean of 0.0075 and a wide range from -0.0549 to 0.0647. Look at the spread rather than a single central bar: there are both sizable negative outcomes (down ~5.49%) and sizable positives (up ~6.47%), so events are mixed. That mix undercuts the simple story that long green streaks reliably lead to short-term pullbacks.

Mean forward 3-day return: events vs baseline
What this chart says

The two bars are almost identical: up-streak events mean 0.0075 versus baseline 0.0074, an edge of only 0.000089. With just 21 event observations against 749 baseline days and a Welch t ≈ 0.012 (p = 0.9904), the tiny difference is nowhere near statistically meaningful. In plain terms, after a long green run the 3-day return looks the same as any random day in this window.

Mean forward 3-day return by streak length bucket
What this chart says

Breaking events by streak length gives no consistent pullback pattern: 5-day streaks average 0.0096, 6-day streaks average -0.012, and 7+ day streaks average 0.0201. The means bounce around (one negative, one strongly positive) rather than trending downward as streak length increases. Given the total of 21 events, these bucket means are noisy, which supports the conclusion that longer streaks don’t produce a reliable short-term mean reversion.

Per-bucket summary (forward 3-day returns)

bucketNmeanmedianstd
5130.00960.01630.03
64-0.012-0.02830.0425
7+40.02010.01260.0317

Event-level anchors (first 50 rows)

datestreak_lenfwd_3d_return
2023-11-0750.0523
2023-11-0860.0504
2023-11-0970.0647
2023-11-1080.0083
2023-11-1390.017
2023-11-1410-0.0096
2024-01-1050.0371
2024-01-2450.0269
2024-03-255-0.0549
2024-05-2450.0229
2024-05-286-0.0418
2024-08-1650.0343
2024-08-196-0.0355
2024-10-0850.0103
2025-06-275-0.0021
2025-09-125-0.0348
2025-10-015-0.0071
2025-10-026-0.0211
2025-10-295-0.0048
2026-04-0750.0279
2026-04-1550.0163

The takeaway

Short answer: no — five-plus day green streaks in NVDA do not leave the stock measurably "due" for a pullback over the next three trading days. The mean 3-day return after those 21 events was +0.75% (N=21) versus +0.74% for the 749-day unconditional baseline; medians were +1.03% vs +0.62% and positive-return rates were 57% vs 55%. A formal Welch test gives t≈0.012 and a two-sided p≈0.990, i.e., there is virtually no statistical difference and about a 99-in-100 chance the tiny observed gap is compatible with noise — and the event sample is small. The bucket breakdown (5-day mean +0.96%, 6-day −1.20%, 7+ day +2.01%, with Ns 13/4/4) shows wild swings but those subgroups are tiny and noisy. Practical takeaway: these three-year numbers support the thesis that the ‘‘overdue for a dip’’ reflex is not showing up as a reliable, tradeable mean-reversion edge on a 3-day horizon.

The fine print