Does any low-timeframe crypto strategy actually work?
Short answer: one does, on and off. It reads funding rather than price shape, it only works on BTC, it stops working one rung below where we found it — and its edge comes and goes with the regime: quiet from 2023 to mid-2025, +8.5% over the last twelve months. Everything else we could express on candles and funding — every momentum, mean-reversion, breakout, squeeze, session and volume family — was pre-registered, tested across two independent windows, and killed. So were three strategies built on open interest, once that history turned out to be free, two built on aggressor flow — who crossed the spread — and three on the last free channels: the perp's discount to spot, options-implied volatility and Bitfinex's margin shorts. So was the classic multi-timeframe setup: buy when RSI pulls back to 50 while the higher timeframe trends up.
This page collects ten sweeps that asked the same question from different angles. The chronological versions, with every hypothesis and neighbour grid, are in the research log.
Why the question is hard to answer honestly
Almost any low-timeframe strategy can be made to look profitable. There are three easy ways to get there and all three are common:
- Test one window. Crypto had a high-volatility era (roughly 2018–2022) and a quieter one after. A strategy born in the first will usually flatter you if that's all you test.
- Price it wrong. A 500-trade backtest that charges nothing per trade, or charges spot fees on a perpetual position, is answering a different question than the one you're asking.
- Keep the best trade. One 40% winner can carry a curve that is otherwise flat.
So every result below had to clear the same gates, fixed before any run: net positive in both independent windows (W1 Aug 2018 → Aug 2022, W2 Aug 2022 → Aug 2026), still positive after deleting the single best trade, max drawdown ≤ 20 (summed-return points, marked at trade exits — the log's protocol note says exactly what that measures), and passing on at least two of the three majors (BTC, ETH, BNB). Funding is charged from real exchange history at every 8-hour settlement a position is held through. Fees moved during the run of these sweeps: the first ones charged 0.1% per side, the library's conservative spot-taker default, and sweep #4 corrected that to the USDT-M perpetual schedule of 0.05% taker and 0.02% maker. Where a verdict turns on that difference, it is flagged below.
The clearest single result: one spec, three timeframes
The cleanest way to see what timeframe does to an edge is to hold the strategy constant and move only the bar size. The funding-squeeze spec — enter long when the funding rate is negative and RSI(14) crosses above 30, exit on a 2×ATR(14) trailing stop — was run at 4h, 1h and 15m with no parameter search in between.
| Timeframe | Result |
|---|---|
| 4h | Passes cross-asset. This is the tradable form, and ships as entry #1. |
| 1h | BTC only. W1 +18.2% · PF 1.39 · DD 17.0% · survives strip-best; W2 +13.1% · PF 1.36 · DD 11.6% · survives strip-best; 122 trades. ETH and BNB both fail. Ships as entry #5 with the BTC-only condition stated as a guardrail. |
| 15m | Dead. BTC negative in both windows (PF 0.72 / 0.86, −22.9% / −10.3%), ETH negative in both, BNB's one positive window followed by a flat one that fails strip-best. |
The per-trade edge shrinks as the bar shrinks. The per-trade cost does not. That sentence is the whole finding, and the rest of this page is the same result arriving from five other directions.
One qualification arrived in October 2026. Sweep #14 ran the 4h form across 2020–22 for the first time, on perpetual candles: BTC lost 17.5% from April 2020 to April 2023, all of it in the 2022 bear market, with a 45-point drawdown. BNB held up. The 4h form passes in the windows it was validated on, not across the whole cycle.
Everything else, and how it died
| Mechanism family | Timeframe | Verdict |
|---|---|---|
| Donchian breakout (close > highest of prior 55, ATR trail) | 1h | PF 0.74–0.92, −31% to −102% per window, 300–500 trades. Fees and chop churn dominate. The 20-period neighbour is worse; 100 doesn't save it. |
| Donchian breakout | 4h | A real edge — all six asset × window cells positive, all six survive strip-best — disqualified on 26–76% drawdown against the ≤20% bar. See the caveat below. |
| Volatility squeeze breakout | 1h | BTC W1 looked spectacular (PF 3.73, +51.8%, DD 6.3%); W2 collapsed (PF 0.53, −37.3%). Same spec, adjacent windows, opposite result. |
| RSI(2) dip-buy (entry #2's spec verbatim) | 1h | PF 0.51–1.19, −40% to −140% per window, ~750 trades each. |
| Short mirror of the breakout | 1h | PF 0.56–0.88, −46% to −166% everywhere. |
| US-open opening-range breakout | 15m | Killed at costs — and it's the most interesting kill in the set. Full write-up. |
| Session-VWAP reversion | 15m | Negative in 11 of 12 cells on the committed re-run — all six at 15m; at the 1h probe BNB's first window is +11.1% and fails on drawdown. |
| London breakout of the Asia range | 15m | W1/W2 regime flip, even before costs. |
| Passive limit mean-reversion (maker fills) | 1h | W1 pays handsomely — BNB +179.5%, ETH +112.1%, BTC +59.0%, all strip-best-surviving. W2 flips BTC/ETH negative. Regime-fragile, not cost-dead. |
| Capitulation-volume reversal (volume > 3× its 20-bar average, close below the lower Bollinger) | 1h | PF 0.62–0.75, −50% to −129% on BTC and ETH. An extreme-volume flush bar is continuation, not exhaustion. |
| Taker-selling exhaustion (the day's taker-buy share in its 28-day bottom decile, then RSI(14) crosses above 30) | 1h | Negative on BTC and ETH in both windows — on ETH worse than random entry (5.5th and 10th percentiles); BNB flips. At 4h it passes on BNB alone and beats random entry in only one window. |
| RSI(2) dip-buy (entry #2's spec verbatim) | 4h | Positive in 2020–23 on ETH and BNB at 33-point drawdowns, negative on all three in 2023–26 — the 1h result on a bar four times longer. |
| RSI(14) oversold turn above the 200-bar average | 4h | 8–15 trades per window and asset; passes on none. |
| Perp discount (premium index in its 28-day bottom decile, then RSI(14)'s turn) | 1h | Negative on BTC and ETH in both windows, ETH worse than random entry; BNB flips. Worse than the funding signal it is built from. |
| Options panic (DVOL in its 28-day top decile, then RSI(14)'s turn) | 4h | The near miss: positive in all four BTC/ETH cells, but single trades carry three of them and 2021–24 draws down 28–31 points. |
| Whale shorts (Bitfinex margin shorts in their 28-day top decile, then RSI(14)'s turn) | 4h | BTC passes 2024–26 only; ETH negative in both windows. |
| RSI(14) back to 50 inside a higher-timeframe uptrend (EMA(20) over EMA(50) on the 4h chart for 1h entries, the daily for 4h; bought at the touch or on the turn back up) | 1h, 4h | No signal on any asset in both windows, and every cell over the drawdown bar (21–96 points). At 1h it churns, 449–627 trades a window. At 4h its three best cells are all from 2020–23 (up to +111.5%); in 2023–26 three of the four BTC and ETH cells lost. |
That covers every data channel the strategy language can read below the daily bar: price shape, volatility, session structure, execution model, volume, and market internals. A later sweep added a channel the language can't read yet — open interest — and it died too, and so did a gate built from aggressor flow a sweep after that; see what would change the answer.
The nearest miss
The 4h Donchian breakout deserves its own paragraph, because it is the one family where the edge is not in question. With the pre-registered 3×ATR trail, every asset × window cell is net positive and every one survives strip-best: BTC +87.4% / +37.7%, ETH +124.9% / +36.2%, BNB +173.6% / +59.3%.
It fails on drawdown — 26% to 76% against a ≤20% bar. That is structural rather than fixable: an always-in breakout system on crypto wins about 35% of the time and pays for its winners with long losing streaks through chop. Adding filters until the drawdown looks acceptable is precisely the curve-fit the vetting bar exists to prevent, so the family got exactly one pre-registered revisit — a daily SMA-200 trend filter. It worked as theorised: drawdowns roughly halved, and BTC's second window came in under the bar. W1 still breached at 30–64. The revisit clause is now spent.
If this ever ships, it ships as an explicitly high-drawdown, sleeve-sized, paper-first entry — a decision made deliberately, not by quietly relaxing the bar.
Why the cost line is the whole story
Two of these kills came with a diagnostic that explains the rest. Running the opening-range breakout with fees switched off leaves a genuine, window-stable tendency worth about +0.04% per trade. At the perpetual taker rate that is a 0.1% round trip, so the structure is real and about two and a half times smaller than the price of trading it. The original sweep charged spot fees and put the gap at five times; the full arithmetic carries both numbers.
The same test on the most liquid instruments in the world found less, not more. Intraday VWAP reversion on SPY and QQQ, run on real regular-hours bars at 2 bp per side — about the friendliest honest cost model retail can claim — looked in August like a gross edge of +0.02–0.04% per trade. The October re-run found that reading had held positions overnight until the VWAP reclaim; flattened at the close, the cell is negative even with fees off, on 500–1,000 trades per window. On the index ETFs there is no intraday structure to pay for.
This is why "trade a lower timeframe for more opportunities" is backwards. More bars means more signals, smaller moves per signal, and an unchanged cost per signal. The number of trades goes up and the amount you keep goes down.
What would change the answer
The honest boundary of this result: it applies to strategies expressible on OHLCV plus funding, at retail execution costs, on major crypto perpetuals. Three things would genuinely reopen it, and none of them is another backtest:
- Order-flow data. Historical order books, open interest and liquidation feeds are where sub-hourly edges plausibly live. One of the three turned out to be free: Binance's public archive carries five-minute open interest back to 2020–21, and sweep #12 tested it in October 2026 — a deleveraging gate, a short build-up gate and a crowd-positioning gate on the funding-squeeze trigger, at 1h. All three were killed. The deleveraging gate lost 21–26% on BTC and ETH in 2022–24 (worse than random entry) and made 11–14% in 2024–26; the other two had no signal or almost no trades. Liquidation history is still not freely available, and order-book depth is archived only from 2023 — too short for two independent windows until about 2028. Aggressor flow — the share of volume bought by takers, which every kline records — was tested in sweep #14 as a gate on the same trigger at 1h and 4h, and killed too. Sweep #15 then tried the last free channels behind the same trigger — the perp's discount to spot, Deribit's options-implied volatility, Bitfinex's margin shorts — and killed all three. Every gated oversold turn on BTC that held up after 2023 broke in 2022; no free positioning read tells a bear market's oversold turn from a bottom. What is left is paid data.
- Maker execution. The passive-fill sweep showed that resting limit orders do capture the gross edge — BNB +179.5% in W1 — before the second window flipped it. A genuine maker-rebate cost structure is a different product than this library models.
- A market-structure change. Fee schedules move, and the 2026 perpetual taker rate is already half what the first sweeps assumed. That correction alone rescued the 1h funding-squeeze from the kill pile.
Absent one of those, the sub-daily floor for this library is 1h — measured, not assumed.
Reproducing this
Every sweep on this page was re-run from committed code on 9 October 2026. The August runs were scratchpad scripts that were never checked in, so for a month these numbers could be read but not re-derived. The replication scripts (scripts/research/replicate-sweep*.ts) rebuild each sweep from the specification frozen in the research log, run it on the same engine over the same public data, and print every published number beside the re-run. Across sweeps #1–#7, 84 published cells reproduce to their printed precision or within rounding; three implementation conventions the log had not written down were recovered afterwards and are disclosed in the log's replication section, together with the sentences the re-run corrected — among them that 41, not 42, of sweep #2's cells were negative, and that the equity VWAP cell quoted above had held positions overnight. The outputs, with every trade, are in scripts/research/results/replication-2026-10/.
What did survive
Seven strategies cleared the bar. Two of them trade crypto below the daily bar — the funding squeeze at 4h and at 1h — and one more trades below it outside crypto, a 5-minute forex calendar trade (entry #7). The strategy library has all of them with full evidence attached. If you came here looking for something to trade on low timeframes, the honest recommendation is entry #5, read together with its conditions: BTC only, real futures fee tier, and a mechanism that reads what other traders are paying rather than what the last few candles looked like — and with the newest evidence on it. Re-run in sweep #12, its May 2024 → Oct 2026 window made +0.9% on 46 trades, failed strip-best and sat at the 74.8th percentile of random entry. Inside that window it lost through mid-2025 and then made +8.5% over the last twelve months (93.8th percentile — just short of the bar). An edge that comes and goes with the regime: paper-trade it before you trust it.