Bitcoin returns after a 30% drawdown
Forward returns from the first daily close 30% below the trailing all-time high.
median forward return
Context before conviction.
After removing overlapping forward windows, N=5 independent observations remain from 7 raw events. Their 180D median return was +47.77% with a 60% win rate and -30.66% median adverse excursion.
Outcome by horizon
Primary non-overlapping results are shown beside the complete raw event population.
| Horizon | Population | Sample | Median | Mean | Win rate | 25th percentile | 75th percentile | Median MAE |
|---|---|---|---|---|---|---|---|---|
| 30D | Independent | N=7 | +2.1% | +24.4% | 57.1% | -11.6% | +42.8% | -10.5% |
| 30D | Raw events | N=7 | +2.1% | +24.4% | 57.1% | -11.6% | +42.8% | -10.5% |
| 90D | Independent | N=6 | -4.3% | +22.9% | 50% | -21.5% | +24.2% | -30.4% |
| 90D | Raw events | N=7 | +7.9% | +78.7% | 57.1% | -19.8% | +108.0% | -30.2% |
| 180D | Independent | N=5 | +47.8% | +129.7% | 60% | -11.3% | +70.5% | -30.7% |
| 180D | Raw events | N=7 | +47.8% | +110.8% | 57.1% | -23.7% | +126.0% | -30.7% |
| 365D | Independent | N=3 | +99.3% | +157.1% | 66.7% | +34.5% | +250.8% | -41.9% |
| 365D | Raw events | N=6 | +34.5% | +93.5% | 50% | -55.3% | +192.2% | -54.5% |
Distribution and drawdown
Inspect the stored event-level outcomes behind the aggregate table.
Temporal validation
Chronological 70/30 holdouts and expanding-window folds use only independent observations. Later outcomes never influence event selection or thresholds.
| Horizon | Independent N | Train median | Test median | Direction | Stable folds | Assessment |
|---|---|---|---|---|---|---|
| 30D | N=7 | — | — | — | — | Insufficient Data |
| 90D | N=6 | — | — | — | — | Insufficient Data |
| 180D | N=5 | — | — | — | — | Insufficient Data |
| 365D | N=3 | — | — | — | — | Insufficient Data |
Sensitivity analysis
One detector setting changes at a time around the published configuration. Returns never determine which configuration is canonical.
At least 8 independent observations are required for the canonical and compared variants.
| Configuration | Parameters | Raw N | Independent N | Median | Win rate | Direction |
|---|---|---|---|---|---|---|
| 20% drawdown | Thresholdpct 20 | N=17 | N=9 | +30.8% | 66.7% | Stable |
| 30% drawdownPublished | Thresholdpct 30 | N=7 | N=5 | +47.8% | 60% | Baseline |
| 40% drawdown | Thresholdpct 40 | N=4 | N=4 | -24.7% | 25% | Too small |
This is a robustness check, not model selection. Each row changes one assumption while every other detector rule remains fixed.
Cross-exchange benchmark
The published detector is replayed independently on each venue over the shared completed-candle window. Exchange populations are never mixed.
Both venues need at least 8 independent outcomes before direction is compared.
Coinbase remains the canonical research source. Bitstamp is an external robustness check; its candles, events, and forward-return population remain separate.
Drawdown crosses below −30% · one event per ATH cycle
The signal is evaluated only with information available at the event close. Future observations are used solely to calculate outcomes.
Drawdown is measured from the highest prior daily close.
Only the first −30% crossing in an all-time-high cycle is counted.
A new cycle begins after a new all-time closing high.
Forward returns do not assume staged entries.
Primary confidence and headline statistics use the earliest chronological non-overlapping sample for each horizon; raw events remain reported for audit.
Temporal validation uses the independent sample only: a chronological 70/30 holdout plus expanding-window folds with at least two later observations per fold and no threshold tuning.
Sensitivity analysis changes one detector parameter at a time around the published canonical configuration, compares independent 180D outcomes, requires at least 8 observations per comparable variant, and never selects a parameter from performance.
Cross-exchange robustness replays the canonical detector independently on Coinbase BTC-USD and Bitstamp BTC/USD over their shared completed-daily history. Each source keeps its own signal and non-overlapping outcome population, with at least 8 observations required per benchmark.
Research, not a recommendation. Live market data can be delayed or temporarily unavailable. Historical results are displayed only when they have been computed and stored by the database research pipeline.