MARA 5-day forward returns after >8% down-days vs baseline (last ~3 years)
A simple look at averages tempts you to call it a win: the 46 instances of MARA’s single-day crashes (≤ −8%) produced a mean 5-day return of +3.31% versus an unconditional baseline of +0.80%. But that apparent +2.50 percentage-point edge does not hold up — medians sit near zero and exactly half the events finished positive, and a Welch t-test (t = 0.914, p = 0.3652) says the difference is statistically indistinguishable from noise.
This study uses minute bars resampled to daily closes across a ~3-year (752-day) window, computing close-to-close 5-trading-day forward returns anchored at the close. The short conclusion: some washouts bounce hard, others keep bleeding, and on average there’s no dependable short-term edge to buying MARA at the close after an >8% down-day. Full dataset, charts, and tests follow below.
For MARA over the past ~3 years, does buying the session after an >8% single-day crash actually pay off — are forward 5-day returns following those washouts meaningfully above the everyday baseline, or is 'buy the dip' just catching a falling knife? Thesis: forward 5-day returns after >8% down-days come out statistically indistinguishable from baseline, so the crashes neither reliably bounce nor keep bleeding and dip-buying the miner carries no real edge.
How this was measured
Minute bars for MARA were resampled to daily closes, then close-to-close daily returns and 5-trading-day forward returns were computed. Event days are those with a same-day return ≤ −8%; for each such day, the forward return is defined as close[t+5]/close[t] − 1, anchoring strictly at the close (no intraday leakage). The everyday baseline is the unconditional distribution of 5-day forward returns across all trading days with a full 5-day window in the same ~3-year range. A Welch two-sample t-test (unequal variances) compares event vs baseline means.
The key numbers
Reading the numbers
Across 2023-06-30 to 2026-06-30 (752 days) there were 46 >8% down-day events. Their mean 5-day return was +3.31% vs baseline +0.80% (edge +2.50 ppt), but Welch p=0.3652 says that uplift is not statistically reliable.
The charts
This histogram shows the 46 post-crash 5-day outcomes are widely spread: returns range from -0.274 to 0.6737 with a mean of 0.0331. Notice the cluster of outcomes around zero and a few large positive outliers (up to 0.6737) as well as meaningful downside (down to -0.274). In plain terms, buying after a >8% drop produced some big winners but also big losers — half the events were positive — so the average +3.31% masks a hit-or-miss pattern rather than a reliable bounce.
The unconditional baseline histogram (n=747) also has a wide spread: 5-day returns run from -0.4204 to 0.722 with a mean near zero (0.008). Look at the bulk of observations sitting close to zero here as well; extreme moves exist in the baseline too. That similarity in tail behavior means the post-crash sample is not uniquely prone to sustained rebounds or collapses compared with ordinary days.
The bar chart puts the simple means side-by-side: events average 0.0331 (3.31%) vs baseline 0.008 (0.80%), an apparent edge of about 0.0250 (2.50 percentage points). But the comparison relies on 46 event observations versus 747 baseline ones, and the formal test (Welch t = 0.914, p = 0.3652) indicates the difference is plausibly noise. So the visual bump in the event bar does not translate into a statistically clear, dependable edge for dip-buying.
Event-level summary (most recent 50)
| event_date | same_day_return | fwd_5d_return |
|---|---|---|
| 2026-06-05 | -0.1187 | 0.1568 |
| 2026-03-06 | -0.0869 | 0.1652 |
| 2026-02-05 | -0.1968 | 0.0968 |
| 2026-02-04 | -0.0815 | -0.0826 |
| 2026-01-20 | -0.0942 | 0.0116 |
| 2025-11-20 | -0.1118 | 0.1491 |
| 2025-11-13 | -0.1032 | -0.1992 |
| 2025-10-16 | -0.12 | -0.0278 |
| 2025-10-10 | -0.0902 | 0.0688 |
| 2025-09-25 | -0.0927 | 0.1659 |
| 2025-07-23 | -0.117 | -0.0544 |
| 2025-05-05 | -0.0927 | 0.2245 |
| 2025-04-08 | -0.1138 | 0.2144 |
| 2025-03-28 | -0.0924 | -0.1027 |
| 2025-03-10 | -0.1842 | -0.0061 |
| 2025-02-24 | -0.0819 | 0.0037 |
| 2025-01-27 | -0.0813 | -0.0093 |
| 2024-12-18 | -0.1086 | -0.1158 |
| 2024-11-18 | -0.1368 | 0.4068 |
| 2024-11-12 | -0.1008 | -0.1243 |
| 2024-10-01 | -0.0836 | 0.066 |
| 2024-09-03 | -0.0932 | 0.0448 |
| 2024-08-12 | -0.1046 | 0.0814 |
| 2024-08-01 | -0.0871 | -0.0144 |
| 2024-07-23 | -0.0897 | -0.0736 |
| 2024-05-10 | -0.1219 | 0.1446 |
| 2024-04-30 | -0.1162 | 0.2569 |
| 2024-04-12 | -0.095 | 0.0528 |
| 2024-04-02 | -0.0923 | -0.088 |
| 2024-03-11 | -0.0952 | -0.0846 |
| 2024-03-05 | -0.1186 | -0.0963 |
| 2024-02-29 | -0.0984 | -0.1709 |
| 2024-02-20 | -0.0814 | 0.242 |
| 2024-02-15 | -0.0909 | -0.1502 |
| 2024-02-05 | -0.0953 | 0.6737 |
| 2024-01-12 | -0.1659 | -0.1247 |
| 2024-01-11 | -0.184 | -0.274 |
| 2024-01-05 | -0.1048 | -0.2195 |
| 2023-12-29 | -0.1599 | 0.0985 |
| 2023-12-28 | -0.1213 | -0.1459 |
| 2023-12-11 | -0.1227 | 0.3584 |
| 2023-10-03 | -0.1291 | 0.1524 |
| 2023-09-08 | -0.1411 | -0.1037 |
| 2023-08-31 | -0.0804 | -0.1259 |
| 2023-08-24 | -0.1135 | 0.1742 |
| 2023-08-17 | -0.1108 | -0.0945 |
The takeaway
Short answer: buying MARA at the close after a single-day drop of 8%+ does not show a reliable short-term edge — 5-day forward returns after those washouts are statistically indistinguishable from the everyday baseline. Across 46 event days the mean 5-day return was +3.31% versus an unconditional baseline mean of +0.80% (an apparent edge of +2.50 percentage points), but medians sit near zero (event median -0.12%, baseline median -0.49%) and only 50% of events finish positive. The formal test is weak: Welch t = 0.914 with p = 0.3652 — roughly a 37-in-100 chance this difference is just random noise, so this is not a statistically clear signal. Practically, some crash days produce big bounces (examples of +15.7% and +16.5% exist) and others keep bleeding; on average over the past ~3 years there’s no consistent pop to rely on, so dip-buying MARA after an >8% down-day offers no dependable edge.
The fine print
- t+5 forward windows overlap heavily and serial dependence may inflate effective sample size; the Welch t-test doesn't correct for that.
- The baseline pool includes the event days themselves, so the comparison is not to an event-free control.
- Results cover ~36 months (2023-06-30 to 2026-06-30); different historical regimes could behave differently.
- Intraday crash/rebound dynamics and MARA's co-movement with BTC are not controlled for and may mask conditional effects.