Does the 5-Minute Edge Survive at Higher Timeframes?
A multi-resolution look at 30 days of real BTC data — 8,639 five-minute windows aggregated to 15 min, 1 hour, 4 hours and daily.
In a previous piece we looked at six months of 5-minute Bitcoin candles and found something uncomfortable for anyone selling a "5-minute strategy": 49.8% of windows closed up. A coin flip. Direction persistence was real but small (57.4%), nowhere near enough to overcome costs on its own.
A fair question came back several times: okay, but what about longer candles? Maybe the noise averages out and an actual edge appears at 15 minutes, an hour, or a day.
That's a testable question. So we re-ran the same dataset across five resolutions — 5m, 15m, 1h, 4h, and 1d — using one rolling 30-day window of Binance BTC/USDT klines (UTC 2026-08-09 to 2026-09-08, 8,639 completed 5-minute buckets). Longer buckets are built by aggregating the raw 5-minute candles, so every row comes from the same underlying prints.
The setup
For each resolution we measured:
- P(up) — share of buckets closing at or above their open
- Persistence — probability a bucket points the same direction as the prior bucket
- 3-in-a-row — share of buckets completing three consecutive same-direction moves (the random-walk baseline is 25%)
- σ (bps) — standard deviation of per-bucket percentage moves, in basis points
- Mean |move| — average absolute move, the magnitude a trader actually has to outrun
- Lag-1 autocorrelation — serial dependence in the size/sign of returns
Only fully-populated buckets were counted.
Finding 1: Short-horizon direction stays a coin flip
The upward bias only shows up later — and it isn't timing skill.
| TF | n | P(up) | Persistence | 3-in-a-row | Mean move | σ | Mean |move| |
|---|---|---|---|---|---|---|---|
| 5m | 8,639 | 50.13% | 48.52% | 22.95% | +0.24 bps | 11.78 bps | 7.27 bps |
| 15m | 2,879 | 50.47% | 49.20% | 23.22% | +0.71 bps | 20.34 bps | 12.53 bps |
| 1h | 719 | 52.16% | 47.21% | 20.92% | +2.92 bps | 42.34 bps | 25.08 bps |
| 4h | 179 | 54.19% | 46.63% | 21.47% | +11.53 bps | 86.86 bps | 51.26 bps |
| 1d | 29 | 55.17% | 35.71% | 14.81% | +71.54 bps | 250.33 bps | 169.32 bps |
At five and fifteen minutes you are, statistically, calling heads or tails. P(up) of 50.1–50.5%, persistence in the high-40s, and a three-in-a-row rate below the 25% random baseline — if anything, intraday direction slightly anti-persists.
The higher P(up) at 4h and 1d is real, but it isn't timing skill. Over these 30 days BTC rose from roughly 64,800 to 79,100, about +22%. The longer the bucket, the more that one-directional drift dominates the bucket's close. A daily P(up) of 55% in a strongly trending month is what exposure (beta) looks like — not a signal you could have traded in and out of.
Finding 2: Volatility obeys the √T law almost perfectly intraday
If returns are independent, standard deviation should scale with √time. A 15-minute candle should be √3 ≈ 1.73× as volatile as a 5-minute candle, an hourly candle √12 ≈ 3.46×, and so on.
| TF | Observed σ | √T prediction (from 5m) | Ratio |
|---|---|---|---|
| 15m | 20.34 bps | 20.41 bps | 0.997 |
| 1h | 42.34 bps | 40.82 bps | 1.037 |
| 4h | 86.86 bps | 81.64 bps | 1.064 |
| 1d | 250.33 bps | 199.98 bps | 1.252 |
The intraday fit is striking — 0.997 at 15 minutes. Across minutes and hours, BTC behaves like a textbook random walk in its magnitude. The daily ratio of 1.25 is the first crack: long-horizon moves are ~25% bigger than independence predicts, consistent with trending/volatility clustering. But with only 29 daily observations, treat that as a hint, not a law.
Finding 3: No serial dependence to exploit during the day
Lag-1 autocorrelation of returns is −0.012 at 5m, +0.016 at 15m, +0.022 at 1h, and +0.056 at 4h — all essentially zero. It climbs to +0.16 on daily returns, which sounds like momentum, but n = 29 and the period was a steady climb, so that coefficient is mostly the trend describing itself.
Translation: knowing the last 5-minute (or hourly) candle's direction gives you no measurable edge on the next one during the trading day.
What this means for a 5-minute observer
- The repeatable thing at 5 minutes is volatility, not direction. Magnitude follows a predictable √T law; the sign does not.
- Costs are the wall, not prediction. A mean absolute move of ~7.3 bps every five minutes has to clear spread plus fees plus slippage twice (in and out).
- "Edge" migrates upward as beta, not as timing. The rising P(up) at higher resolutions is drift you get by holding through an up month.
- Long timeframes trend; short timeframes mean-revert to noise. Daily series is the right place to look for genuine persistence — but 29 samples cannot support a tradeable claim.
Caveats — read these before quoting any number
- Short, biased window. 30 days and a +22% up-month. The high-timeframe P(up) and daily autocorrelation are regime-dependent.
- Tiny daily sample. 29 daily buckets cannot support a strong momentum claim.
- No costs modeled. All figures are gross of fees, spread, and slippage — which only strengthens the "hard to trade intraday" conclusion.
- Single venue, single asset. Binance BTC/USDT only; results need not transfer to alts or other exchanges.
Takeaway
The 5-minute edge doesn't "survive" at higher timeframes because, on direction, it was never really there — at any intraday resolution. What does survive cleanly is the volatility structure: √T scaling holds tightly from 5 minutes to 4 hours. Above that, drift and trend take over, and the apparent edge is just exposure to an asset that went up that month.
Separate the two questions. How much will it move? is largely predictable. Which way? — over minutes and hours — is still a coin you have no business thinking you've rigged.
Watch these distributions move live
BTCbeat tracks up/down balance, trend persistence, volatility scaling and alignment against the official Chainlink TWAP-60s settlement feed — updated from the same Binance klines every 5 minutes.
Open the live BTC pulse dashboard →