AI Research COIN

COIN extreme daily moves: next-5-session continuation vs reversal

We tested whether Coinbase’s largest single-session moves — the top and bottom deciles of daily returns over the last ~36 months — tend to continue or reverse over the next five trading days. Extreme sessions were identified in-sample and forward compounded returns for k=1..5 days were measured; the full methodology, return profiles, and t-tests are below.

The working thesis was that COIN’s biggest down-days mean-revert (produce above‑baseline bounces) while big up-days fade. The data do not support a clear asymmetry: worst-decile days (cutoff −5.3209%) averaged a +2.39% return over the next five sessions, but statistical tests show the tails are indistinguishable (p≈0.966 bottom vs top). Scroll down for the complete statistics, charts, and significance analysis.

The research question

For COIN over the past ~3 years, when it posts an extreme single-session move (top or bottom decile of daily returns), do the next five sessions continue that move or reverse it? Thesis: COIN's biggest down-days are followed by above-baseline bounces while its biggest up-days fade, so the daily extremes mean-revert rather than trend — punishing the chase.

How this was measured

Resampled COIN minute bars to daily closes and computed close-to-close returns over the last ~36 months (2023-05-31 to 2026-05-29). Identified extreme sessions as bottom-decile (worst 10%) and top-decile (best 10%) by daily return computed in-sample. For each extreme day, measured compounded forward returns over the next k=1..5 trading days (close[t+k]/close[t]−1) and compared to the unconditional baseline across all days with valid forward windows. Reported mean forward-return profiles, reversal fractions (e.g., bounce after down-day = positive forward return), and Welch t-tests (unequal variance) for 5-day forwards vs baseline.

The key numbers

Trading days analyzed
751
2023-05-31 to 2026-05-29
Bottom-decile cutoff (daily return)
-5.3209%
Top-decile cutoff (daily return)
5.9050%
Bottom-decile days (N)
76
Top-decile days (N)
76
Mean +1d after bottom-decile
0.5269%
Mean +5d after bottom-decile
2.3902%
Mean +1d after top-decile
0.6048%
Mean +5d after top-decile
2.3082%
Baseline mean +5d
1.3760%
Edge +5d (bottom − baseline)
1.0142%
Edge +5d (top − baseline)
0.9321%
Reversal fraction +1d (after bottom-decile)
57.8947%
Bounce rate (fwd +1d > 0)
Reversal fraction +5d (after bottom-decile)
51.3158%
Bounce rate (fwd +5d > 0)
Reversal fraction +1d (after top-decile)
48.6842%
Fade rate (fwd +1d < 0)
Reversal fraction +5d (after top-decile)
51.3158%
Fade rate (fwd +5d < 0)
Welch p-value (+5d bottom vs baseline)
0.4474
p=0.4474 ≥ 0.05 → no clear difference
Welch p-value (+5d top vs baseline)
0.5450
p=0.5450 ≥ 0.05 → no clear difference
Welch p-value (+5d bottom vs top)
0.9663
p=0.9663 ≥ 0.05 → no clear asymmetry

Reading the numbers

Across 751 trading days the bottom-decile cutoff was -0.05320941355508457 and the top-decile cutoff was 0.05905006418485237. Both bottom- and top-decile days average about 0.0239 and 0.0231 over the next five days versus a baseline 0.013760448911026174.

The charts

Mean forward return by horizon
What this chart says

This bar chart plots mean forward returns from +1d to +5d after bottom-decile, top-decile, and baseline days. Focus on the +5d bars: after bottom-decile the mean is 0.0239 and after top-decile it is 0.0231, both above the baseline 0.0138; the +1d results are small positives (0.0053 bottom, 0.0060 top vs 0.0027 baseline). In short, big down-days show a steady recovery across horizons and big up-days do not show clear continuation — both extremes end up with similar above-baseline five-day gains.

Five-day forward return distribution
What this chart says

The box plots show the full +5d return distributions (n=76 for bottom and top deciles, n=746 baseline). Look at the extremes and means: bottom min -0.1981 max 0.4044 mean 0.0239; top min -0.1645 max 0.554 mean 0.0231; baseline min -0.2622 max 0.8326 mean 0.0138 — both deciles have big positive outliers and wide spread. That tells us the average five-day bounce after big down-days exists, but outcomes are noisy and distributions overlap a lot, so the mean-reversion signal is real on average but far from guaranteed for any single event.

Bucket summary (means and reversal rates)

bucketNmean_+1dmean_+5dreversal_frac_+1dreversal_frac_+5d
Bottom decile760.00530.02390.57890.5132
Top decile760.0060.02310.48680.5132
Baseline7500.00270.0138

The takeaway

Short answer: no clear asymmetric mean-reversion — COIN's biggest down-days do not produce reliably stronger bounces than big up-days produce fades. Over the 751 trading days, the 76 worst days averaged +2.39% over the next 5 sessions versus a baseline +5d gain of +1.38% (an edge of about +1.01 percentage points); the 76 best days averaged +2.31% (+0.93 pp edge). The single-day bounce rate after big down-days was 57.9% at +1d (but only 51.3% at +5d); big up-days faded 48.7% of the time at +1d and also show 51.3% positive at +5d. Stat tests give p=0.447 (bottom vs baseline), p=0.545 (top vs baseline) and p=0.966 (bottom vs top) — in plain terms these are weak, non‑significant results (roughly 45–55% chances the observed edges are just noise, and ~97% that there is no asymmetry bottom vs top). You only have 76 extreme triggers per tail (baseline ~750 days), so evidence is thin. Practical takeaway: don’t assume chasing COIN’s big down-days will reliably buy a bounce, nor that big up-days will reliably fade — the observed short-term moves are small and statistically inconclusive.

The fine print