Do Breakouts Work in Crypto? — Bybit Replication

1 Headline Findings

Upside breakouts (20-day high close)

—

mean next-day return

Events—

Win rate—

Avg winner—

Avg loser—

Payoff ratio—

Downside breakdowns (20-day low close)

—

mean next-day return

Events—

Win rate—

Avg winner—

Avg loser—

Payoff ratio—

In this cached sample, upside breakouts have a positive mean despite a sub-50% win rate because average winners exceed average losers. This descriptive result does not establish an out-of-sample edge or account for trading costs.

2 Bybit vs Binance (Dimaquant)

Metric	Dimaquant (Binance 2021–25)	This replication (Bybit)
Upside mean (next-day)	+15.4 bps (outperformance vs universe)	—
Downside mean (next-day)	−11.7 bps (raw ≈ flat)	—
Upside win rate	49.2%	—
Downside win rate	≈ 50%	—
Upside payoff ratio	1.19 (1.22 ex-outliers)	—

Reading the comparison. Dimaquant's +15.4 bps is outperformance vs the top-40 universe average, not a raw return. This Bybit analysis reports raw returns from its available cache, so its mean cannot be compared directly with the Binance outperformance figure. The shared pattern is descriptive rather than conclusive: upside breakouts have a positive sample mean and payoff asymmetry, while downside breakdowns have a negative sample mean. Outlier and statistical-robustness tests remain future work.

3 Yearly Stability

Upside breakout mean return by year (bps)

Downside breakdown mean return by year (bps)

Yearly means can vary substantially, especially for years with few events. This chart is descriptive: it does not establish a stable, tradeable edge or test whether returns are robust to outliers, costs, or changing market regimes.

4 Per-Symbol Detail (all symbols, ordered by signal count)

Mean next-day returns in bps for every symbol in the universe. days shows how much daily history was available.

5 Method & Replication Notes

Expand methodology

Universe. Top 40 Bybit USDT linear perps by 24h turnover at analysis time, used as a market-cap proxy. Dimaquant rebuilt his top-40 by point-in-time market cap each month; we used a single snapshot (simplification, noted).

Signal. For each day t: high_20 = max(close[t-19..t]), low_20 = min(close[t-19..t]). Upside breakout when close[t] = high_20; downside breakdown when close[t] = low_20 (floating-point tolerance).

Return. Close-to-close next-day return (close[t+1] − close[t]) / close[t].

Costs. Raw returns — no trading fees, slippage, or funding (Dimaquant charged 10 bps one-way turnover in his strategy tests; funding excluded there too).

Data. Bybit public REST API (/v5/market/kline, interval D, up to 1000 days per symbol). Cached locally under ohlcv/. Survivorship: universe is currently-listed symbols — a known bias vs Dimaquant's delisting-adjusted sample.

Metrics. Pooled event-level stats (not symbol-averaged): mean bps, win rate = fraction of events with positive next-day return, payoff = avg winner / |avg loser|.

6 Takeaways for Bybit Perp Trading

Upside signals have a positive in-sample mean (+76 bps pooled), driven by payoff asymmetry: the win rate is about 48% while winners are larger than losers.
Downside signals have a negative in-sample mean. The result is not a strategy test and does not model a short position, fees, slippage, or funding.
Yearly results vary. Small cohorts can make annual averages unstable, so the chart should not be interpreted as evidence of a constant return premium.
Freshness is untested in this repository. A useful next analysis is a freshness decomposition and a 1–5 day holding-period sweep using point-in-time universe membership.
Costs remain unmodeled. Any trading conclusion needs fees, slippage, funding, turnover, and out-of-sample validation.