Study design and coverage
The study evaluates 17,581 distinct Bear Flag market events from one dedicated detection run under one fixed outcome methodology.
The scan produced 19,899 raw Bear Flag events from 309,715,131 evaluated rolling windows. After deterministic measurement and outcome-blind family selection, the analytical dataset contains 17,581 Bear Flag pattern families. Every scan endpoint was reconciled: none failed detection, none remained unresolved, and none overlapped the measured/skipped inventories.
The combined sample represents 59 Binance markets and all 15 configured timeframes from 1m through 1M. Every canonical family had the expected downward endpoint; nine favorable-move rows were right-censored or had an unusable data gap, leaving 17,572 evaluable declines for the main 5% endpoint. Endpoint-specific denominators are shown throughout instead of treating censored rows as misses.
17,581 canonical pattern families
Every family enters the analytical sample once. The grouping rule is fixed before outcomes are counted, so successful patterns are not preferentially selected.
Claim boundary
These are confirmed-breakout observations. They estimate post-breakout outcomes, not how often an unfinished candidate will break out, and they do not include execution costs or position sizing.
Detector run accounting
| Timeframe | Evaluated rolling windows | Raw Bear Flag events |
|---|---|---|
| 1m | 185,575,941 | 7,889 |
| 3m | 61,850,792 | 5,909 |
| 5m | 37,105,099 | 4,617 |
| 15m | 12,360,481 | 465 |
| 30m | 6,174,333 | 243 |
| 1h | 3,081,099 | 351 |
| 2h | 1,534,815 | 188 |
| 4h | 761,522 | 92 |
| 6h | 503,749 | 58 |
| 8h | 374,905 | 35 |
| 12h | 246,003 | 26 |
| 1d | 117,132 | 18 |
| 3d | 22,619 | 8 |
| 1w | 6,641 | 0 |
| 1M | 0 | 0 |
| Total | 309,715,131 | 19,899 |
One window is one eligible scan endpoint under the study's fixed inclusion rules. Totals are reported by the same per-timeframe sample table.
Why lower timeframes contain more observations
Lower timeframes contain more completed candles over the same calendar span, which creates more opportunities for patterns to appear. The sample contains 6,959 families at 1m, 5,193 at 3m, and 4,004 at 5m. Those three strata account for 91.9% of the cohort, so the dedicated 5m+ and 15m+ views are essential context for the all-timeframe headline.
Detection atlas and timeframe sample sheets
More than 100 visually distinct detector outputs, cropped and de-identified for public comparison. Statistical results use the complete analytical cohort, not these selected examples.
Outcome definitions and statistical method
Every Bull-study endpoint is mirrored in the bearish direction; only the favorable direction and downward target projection change.
Favorable decline
The decline from the fixed breakdown boundary to the causally resolved lowest favorable price. The primary Bulkowski-style threshold is 5%.
Structural hold
Whether the confirmed downward endpoint avoids the fixed structural-failure condition during the predeclared observation window. This is not a profit measure.
Downward pole targets
The classical full-pole height and a separately reported 46%-of-pole target are subtracted from the breakdown boundary. Non-evaluable rows are excluded, never converted into misses.
Uncertainty
Binary outcomes report numerator, evaluable denominator, and two-sided 95% Wilson intervals. Continuous declines report mean, median, and quartiles.
Pattern geometry, expected direction, breakdown boundary, and observation contract were fixed before future candles were measured. A pattern could not be redrawn after its outcome was known.
A single analysis workflow was applied to all 19,899 detections. Earliest observation per exchange, symbol, timeframe, and family signature was selected without using future outcome fields, producing 17,581 canonical families.
The primary move endpoint has 17,572 evaluable rows. Structural hold also has 17,572; the full-pole target has 17,550; the 46%-pole target has 14,022; and boundary-proximity pullback has 17,580. Every difference is explicit censoring or endpoint eligibility.
Data governance note
The analytical cohort was processed once under a single study definition, with uniform inclusion rules across all timeframes and no overlap of duplicate event rows in the final published counts.
Pooled outcomes across all measured timeframes
The mirrored Bear study keeps endpoint-specific denominators and reports chart outcomes separately from trade returns.
| Endpoint | Numerator / denominator | Estimate | 95% Wilson interval |
|---|---|---|---|
| Structural hold | 13,830 / 17,572 | 78.70% | 78.09% to 79.30% |
| Favorable decline ≥5% | 2,916 / 17,572 | 16.59% | 16.05% to 17.15% |
| Full-pole target | 4,431 / 17,550 | 25.25% | 24.61% to 25.90% |
| 46%-of-pole target | 10,015 / 14,022 | 71.42% | 70.67% to 72.17% |
| Boundary-proximity pullback | 17,333 / 17,580 | 98.59% | 98.41% to 98.76% |
Structural hold, target attainment, and pullback are descriptive chart endpoints. They are not interchangeable and none includes execution costs.
Practical timeframe aggregates
| Cohort | Families | Structural hold | Decline ≥5% | Full-pole target | 46% target |
|---|---|---|---|---|---|
| 5m+ | 5,429 | 4,088/5,420 (75.42%) | 1,473/5,420 (27.18%) | 1,479/5,417 (27.30%) | 2,972/4,132 (71.93%) |
| 15m+ | 1,425 | 994/1,416 (70.20%) | 630/1,416 (44.49%) | 459/1,422 (32.28%) | 803/1,055 (76.11%) |
8.12%
15m+ mean decline
3.70%
15m+ median decline
0.66%
15m+ 25th percentile
11.04%
15m+ 75th percentile
All-timeframe decline distribution
2.77%
Mean decline
1.11%
Median decline
0.24%
25th percentile
3.41%
75th percentile
The mean decline is about 2.5 times the median, so large downside tails lift the average. The median is a better description of the typical measured event.
Results for all 15 configured timeframes
Every configured timeframe is shown, including zero-detection strata. Sparse higher-timeframe rows are descriptive, not stable rankings.
| Timeframe | Pattern families | Structural hold | Decline ≥5% | Full-pole target |
|---|---|---|---|---|
| 1m | 6,959 | 5,624/6,959 (80.82%) | 616/6,959 (8.85%) | 1,689/6,948 (24.31%) |
| 3m | 5,193 | 4,118/5,193 (79.30%) | 827/5,193 (15.93%) | 1,263/5,185 (24.36%) |
| 5m | 4,004 | 3,094/4,004 (77.27%) | 843/4,004 (21.05%) | 1,020/3,995 (25.53%) |
| 15m | 464 | 338/464 (72.84%) | 161/464 (34.70%) | 152/464 (32.76%) |
| 30m | 241 | 171/241 (70.95%) | 90/241 (37.34%) | 77/239 (32.22%) |
| 1h | 324 | 228/324 (70.37%) | 150/324 (46.30%) | 104/324 (32.10%) |
| 2h | 174 | 105/174 (60.34%) | 75/174 (43.10%) | 55/174 (31.61%) |
| 4h | 86 | 64/86 (74.42%) | 61/84 (72.62%) | 32/86 (37.21%) |
| 6h | 55 | 37/55 (67.27%) | 37/55 (67.27%) | 18/55 (32.73%) |
| 8h | 33 | 24/32 (75.00%) | 24/33 (72.73%) | 12/33 (36.36%) |
| 12h | 24 | 12/23 (52.17%) | 16/24 (66.67%) | 2/24 (8.33%) |
| 1d | 17 | 14/16 (87.50%) | 15/16 (93.75%) | 5/17 (29.41%) |
| 3d | 7 | 1/1 (100.00%) | 1/1 (100.00%) | 2/6 (33.33%) |
| 1w | 0 | N/A | N/A | N/A |
| 1M | 0 | N/A | N/A | N/A |
The detector ran on 1m, 3m, 5m, 15m, 30m, 1h, 2h, 4h, 6h, 8h, 12h, 1d, 3d, 1w, and 1M. It found no canonical Bear Flags at 1w or 1M. Endpoint denominators are separate from the total family count: the seven 3d families include six right-censored observations, leaving only one evaluable favorable move and structural endpoint (1/1). The full-pole target was evaluable for six families (2/6), so none of these cells should be read as a reliable 3d rate.
Why the fraction denominator changes
“Pattern families” is the number detected. The denominator in each fraction is the number for which that particular endpoint was valid and fully observed. Excluded rows are censored or technically unevaluable; they are not silently counted as failures.
| Endpoint | Overall evaluable | Excluded | Why excluded |
|---|---|---|---|
| Structural hold | 17,572 | 9 | Right-censored lifecycle outcomes |
| Decline ≥5% | 17,572 | 9 | Unresolved favorable-move horizon: 4h (2), 1d (1), 3d (6) |
| Full-pole target | 17,550 | 31 | 30 breakout-bar ambiguities plus 1 right-censored 3d row |
| 46%-of-pole target | 14,022 | 3,559 | Invalid projection beyond the boundary or breakout-bar ambiguity |
| Boundary pullback | 17,580 | 1 | One 3d pullback window not fully evaluable |
The answer changes with the decline threshold
The complete 17,572-row favorable-decline curve is more informative than one headline percentage.
Volatility-normalized outcomes
Across all timeframes, 51.92% declined at least 1%, 16.59% declined at least 5%, and 5.44% declined at least 10%. In volatility units, 73.04% reached 1 ATR and 25.00% reached 10 ATR. The curve was reported as sensitivity analysis, not mined to choose a flattering threshold.
Results by BTC market regime
Breakout-day regime splits are descriptive secondary analyses and were not used to select or redraw patterns.
| BTC market regime | Canonical families | Structural hold | Decline ≥5% |
|---|---|---|---|
| Bull market | 9,742 | 7,620/9,742 (78.22%) | 1,664/9,742 (17.08%) |
| Bear market | 7,564 | 5,996/7,555 (79.36%) | 1,177/7,555 (15.58%) |
| Unclassified | 275 | 214/275 (77.82%) | 75/275 (27.27%) |
Regime uses the previous completed BTCUSDT daily close relative to its trailing 200-day moving average; missing history is unclassified. The 5% decline occurred in 17.08% of bull-regime rows and 15.58% of bear-regime rows. These unadjusted splits reflect different symbol and timeframe composition and are not causal regime effects.
Comparison with Bulkowski's stock reference
The directionally aligned comparison uses downward stock flags and the same 5% break-even convention.
| Metric | Bulkowski downward stock flags | All crypto timeframes | Crypto 15m+ |
|---|---|---|---|
| Average favorable decline | 8% | 2.77% | 8.12% |
| Median favorable decline | Not reported | 1.11% | 3.70% |
| Failure below 5% | 45% | 83.41% | 55.51% |
| Success at or above 5% | 55% | 16.59% | 44.49% |
| Full measure-rule target | 46% | 25.25% | 32.28% |
This is a descriptive benchmark, not an apples-to-apples replication of Bulkowski's stock sample. Crypto trades continuously, this cohort is heavily intraday, and the same endpoint convention is documented in our bear flag pattern guide.
Limitations and claim boundary
- 1
Conditional denominator
The headline estimates begin at a confirmed downward endpoint. They do not estimate the probability that an unfinished candidate will break down.
- 2
Timeframe imbalance
1m, 3m, and 5m contribute 91.9% of families. Dedicated 5m+ and 15m+ aggregates prevent that mix from being mistaken for a universal timeframe result.
- 3
Endpoint-specific censoring
Nine move rows, nine structural rows, 31 full-target rows, 3,559 adjusted-target rows, and one pullback row are excluded from their own denominators for explicit evaluability reasons.
- 4
Dependence
Multiple symbols can share the same market-wide shock. Wilson intervals treat observations as independent and may be narrower than cluster-aware intervals.
- 5
Sparse higher timeframes
3d has seven families and one evaluable move; 1w and 1M have none. Those strata cannot support rankings.
- 6
No execution model
Entries, stops, fees, slippage, funding, liquidity, latency, and position sizing are outside this descriptive study.
- 7
Detector scope
The detection specification defines the instrument. A future update to patterns, filters, market universe, or timing rules can produce different estimates.
- 8
No strategy claim
Structural hold, target hit, pullback, and favorable decline are chart endpoints, not expected value, risk-adjusted return, or forward-test evidence.
Questions this study answers
What is the Bear Flag win rate in crypto?
With a post-breakdown decline of at least 5% as the win definition, 2,916 of 17,572 evaluable families succeeded: 16.59%. The rate is 27.18% at 5m+ and 44.49% at 15m+.
How many Bear Flags are in the study?
The scan found 19,899 raw events. Outcome-blind family selection produced 17,581 canonical pattern families; 17,572 have an evaluable favorable-move endpoint.
Were all timeframes tested?
Yes. The run covered 1m, 3m, 5m, 15m, 30m, 1h, 2h, 4h, 6h, 8h, 12h, 1d, 3d, 1w, and 1M. No canonical families were found at 1w or 1M.
What is the 5m+ result?
Among 5,420 evaluable 5m+ families, 27.18% declined at least 5%, 75.42% held structurally, and 27.30% of 5,417 evaluable rows reached the full-pole target.
What is the 15m+ result?
Among 1,416 evaluable 15m+ families, 44.49% declined at least 5% and 70.20% held structurally. The full-pole target rate was 32.28% on 1,422 evaluable rows.
Why is structural hold much higher than the 5% rate?
They answer different questions. Structural hold asks whether bearish chart structure survived; the 5% endpoint requires a specific decline magnitude before the causal stopping event.
Does the 98.59% pullback rate mean an easy trade?
No. It records boundary proximity after breakdown using this study's fixed result definition. It does not include order fills, fees, timing, or profitability.
Was any post-processing language used during publication?
The study output is precomputed. The publication UI only renders prevalidated aggregate evidence.
Conclusion
Across 17,572 evaluable downward endpoints, 16.59% declined at least 5%, 25.25% reached the classical full-pole target, and 78.70% held structurally. The 15m+ decline rate was materially higher at 44.49%, but its cohort is much smaller.
The central lesson mirrors the Bull study: a Bear Flag has no single win rate until the endpoint, denominator, timeframe mix, and observation contract are specified. These are reproducible chart-outcome statistics, not a claim of tradable net returns.
Sources and references
Primary source. The dataset and outcome statistics in this article are original ChartScout research. Pattern identification, duplicate-event grouping, and post-breakdown measurement were designed and run in-house. When citing these figures, attribute them to ChartScout Research (2026).
- ChartScout Research. Bear Flag Win Rate in Crypto: A Study of 17,581 Pattern Families. ChartScout, 2026.
Binance-only study cohort across 59 pairs and 15 configured timeframes, followed by a fixed post-breakout outcome definition. - Bulkowski, Thomas N. Flags (regular): chart pattern statistics. ThePatternSite.com.
Stock-market reference for average decline, break-even failure, and downward measure-rule comparison. thepatternsite.com/flags.html. - Bulkowski, Thomas N. Measurement and pattern-statistics glossary. ThePatternSite.com.
Definitions for break-even failure, average decline, pullback, and downward measure rule. thepatternsite.com/glossary.html. - Lo, Andrew W., Harry Mamaysky, and Jiang Wang. Foundations of Technical Analysis: Computational Algorithms, Statistical Inference, and Empirical Implementation. Journal of Finance 55, no. 4 (2000). NBER Working Paper 7613.
- Sullivan, Ryan, Allan Timmermann, and Halbert White. Data-Snooping, Technical Trading Rule Performance, and the Bootstrap. Journal of Finance 54, no. 5 (1999). DOI.
Stjepan Ivanović
Founder of ChartScout · Crypto Trader Since 2013
Trading crypto since 2013 with his first Bitcoin bought at ~$200. Four complete bull/bear market cycles, traded on early exchanges like Mt.Gox and BTC-e, on-chain trading on IDEX and EtherDelta, and ~70 crypto project investments. Built ChartScout after 22+ months of development to automate what no trader can do manually. Watch hundreds of charts 24/7.