Strategy validation ledger: what we tested and what we found
AETHER tests popular crypto trading ideas on real exchange data before putting them in front of users, and records the outcome whether or not it flatters the idea. Most ideas, including chart patterns, mean reversion, pairs trading, cross-exchange funding arbitrage and the kimchi premium, showed no edge after costs. One family of trend-following breakout rules looked promising on history and has not yet been tested live. This page lists the results, the samples behind them and their caveats.
Last updated:
See how the AI cites this ledger →
Opens the AI research terminal. The assistant quotes these results when you ask about the same ideas.
How we test
- Criteria are fixed before looking at results, so a strategy cannot be tuned until it passes.
- Results are compared with a control that uses the same real prices but random entry times and directions, repeated hundreds or thousands of times.
- When many variants are tried, p-values are corrected for multiple comparisons (Bonferroni).
- Costs are included: round-trip fees, slippage and, where noted, funding. Expectancy is often shown in R, where 1R is the distance to the stop-loss.
- Confidence intervals are shown. An interval that includes zero means the result cannot be told apart from chance.
- A small sample cannot show that an effect does not exist. Where we only had a handful of independent events we say “not confirmed”, not “no effect”.
Verdict labels
- No edge found: Not better than chance, or worse, after costs.
- Limited or inconclusive: Some effect or partial benefit, but it does not meet the criteria for a tradable edge.
- Promising, not live-tested: Met most criteria on history. Live or forward testing has not been done.
Results
Chart patterns and indicator strategies
No edge found · Tested
Idea tested: Indicator strategies such as RSI and Bollinger, and chart patterns such as Gartley and double bottom, make money.
Sample: BTC and ETH on 1h, 4h and daily candles (33 cells), compared with 300 random-entry controls on the same prices, with Bonferroni correction.
What we found
- 0 of 33 cells passed the pre-set criteria. The 5 cells with positive expectancy were all statistically insignificant.
- Patterns (Gartley, double and triple bottom) averaged +0.002R. The 9 indicator strategies averaged -0.119R.
- Best cell: 4h multi-bottom, 138 trades, 56.5% win rate, +0.105R, lower confidence bound -0.185R, adjusted p-value 0.405.
- Mean reversion (RSI plus Bollinger) was the worst, from -0.24R to -0.67R. A 50% win rate still has negative expectancy after round-trip costs and funding.
Caveats
- Cost assumption: 0.15% per side plus 0.01% funding per 8 hours, about twice the assumption used in other checks.
Fading the range on quiet days
No edge found · Tested
Idea tested: On low-volatility days, trading against moves at the edge of the range makes money.
Sample: BTC and ETH 1h candles, 2019-01 to 2026-10, 1,539 quiet-day trades.
What we found
- Net expectancy -0.109R per trade (95% interval -0.150 to -0.068), negative in all 7 years, win rate 45%.
- Before costs the result was -0.053R, the same as a synthetic random-walk baseline (-0.057R): no edge, and costs turn it into a loss.
- Quiet days (-0.109R) were worse than ordinary days (-0.084R).
Caveats
- Funding not included. Maker fills ignore queue position and are optimistic.
Bollinger Bands (20, 2σ) touch and return to the midline
Limited or inconclusive · Tested
Idea tested: After price touches the outer band it is likely to return to the midline, so trading against the touch earns money.
Sample: BTC, ETH, SOL and BNB, about 9,400 touches on 1h and 2,300 on 4h candles (2019 to 2026-10), versus the chance of undoing a same-sized move at any time.
What we found
- Within 12 bars of a touch, about 41% returned to the midline, against about 30% for the control. The touch itself does have a reversion tendency.
- Trading against the touch with a 1:1 stop still lost after costs: -0.137R on 1h (95% interval -0.160 to -0.113) and -0.097R on 4h (-0.145 to -0.051).
- Before costs it was about -0.04R: the gain when price returns is capped at the retracement, while the loss when it keeps going is the same size.
Caveats
- Only a 1:1 stop and target was tested. Funding not included.
Hourly ATR-expansion breakout (trend following)
Promising, not live-tested · Tested
Idea tested: Low win-rate, high payoff trend-following breakout strategies make money.
Sample: BTC and ETH hourly candles, 2020-01 to 2026-10. Eight candidates were fixed in advance and compared with 300 random-entry controls on the same prices.
What we found
- Only one candidate looked promising: the hourly ATR-expansion breakout. 1,338 trades, 31% win rate, payoff ratio 3.07, +0.263R including funding (lower confidence bound +0.058R), positive in all 7 years, best adjusted p-value of the eight (0.020).
- It failed the pre-set drawdown criterion: maximum drawdown was 53 percentage points at 1% risk per trade, 26.6 at 0.5% and 13 at 0.25%.
- It is sensitive to cost: +0.216R at 5 bp slippage per side, +0.136R at 10 bp (interval includes zero), about zero at 20 bp. Break-even is about 18 bp per side.
- Daily candidates had large expectancy but only 69 to 138 trades and could not be told apart from random entry. The 90-day momentum and 24/12-hour breakouts failed.
Caveats
- Not yet tested live or forward on paper.
- The long/short edge flips from year to year. Five other coins combined gave +0.146R but the lower bound was below zero.
- Do not use high leverage: the risk per trade for this structure is limited to 0.25% to 0.5%.
Pairs trading (spread mean reversion)
No edge found · Tested
Idea tested: When two similar coins drift apart, the spread closes again.
Sample: 8 pairs such as ETH/BTC and SOL/ETH.
What we found
- All 8 pairs lost money, averaging -0.56% per trade before costs.
- Wide spreads tended to keep widening: momentum rather than mean reversion.
Caveats
- The other pairs, the sample period and the trade counts were not recorded, and the check script was not kept in the repository.
Cross-exchange funding-rate arbitrage
No edge found · Tested
Idea tested: Going long on one exchange and short on another to collect different funding rates earns a profit.
Sample: Binance, Bybit and Hyperliquid, 365 days of real data, BTC, ETH and SOL.
What we found
- Choosing the most favourable pair at each moment gave annual net returns of BTC -1.8%, ETH +0.3%, SOL -1.4%.
- The best fixed pair (short Hyperliquid, long Bybit) made 2.4% to 4.0% a year gross, before costs and exchange risk.
- Spot-long plus futures-short carry was -1.2% to -5% a year in this window.
Caveats
- Exchange risk such as withdrawal halts and liquidation is not in the numbers. One 365-day window, so other market regimes were not tested.
Delta-neutral funding carry (spot long, futures short)
Limited or inconclusive · Tested
Idea tested: Buying spot and selling futures to collect funding gives a steady return.
Sample: 2019 to 2026, BTC, ETH, SOL and BNB.
What we found
- Holding BTC and ETH all the time averaged about 11% to 14% a year, but it shrank over time. BTC: 30.6% in 2021, 4.2% in 2022, 7.9% in 2023, 12.0% in 2024, 5.1% in 2025, 2.3% so far in 2026.
- SOL and BNB returned about 0%.
- Quarterly-futures basis on BTC and ETH averaged about 7% a year net over 22 expiries, with recent expiries at 1% to 5%.
- With the short leg at 3x, about half of the 90-day windows saw a rise of more than 28%, so margin management is essential. Recent carry is near the yield on cash and is not a profit engine.
Caveats
- The historical average is positive but trending lower toward the present.
Extreme funding rates and price reversals
No edge found · Tested
Idea tested: When funding is extremely high, price reverses downward afterwards.
Sample: BTC, ETH, SOL and XRP. The sample size was not recorded.
What we found
- After extremely high funding, next-day returns were higher, not lower.
- An SMA200 regime filter reduced drawdown on BTC, ETH and SOL but did not raise returns, and failed on XRP.
Caveats
- No basis for using it as a directional signal was found.
Direction forecasts from look-alike past charts (fractals)
No edge found · Tested
Idea tested: Price moves the way it did after similar-looking past shapes.
Sample: Walk-forward test on about 53,000 points across 5 coins, plus BTC, ETH and SOL over 5.8 years.
What we found
- Five-hour direction hit rate was 50.7% against 50.0% for chance (+0.7 points, interval +0.1 to +1.2; the later half, +0.5 points, includes zero).
- Hit rates 5 and 24 bars ahead were 49% to 54%, indistinguishable from chance, and lost money after costs.
- The difference in average 5-hour return between look-alikes that rose and fell was about 0.055%, smaller than trading costs.
Caveats
- Only the Java fallback (Euclidean distance) was tested, not the Python DTW path. The historical win rates shown on screen are visual references, not forecasts.
ONNX crash/spike risk model as a trade veto
No edge found · Tested
Idea tested: A risk model can be used to veto trades in the wrong direction.
Sample: Out-of-sample period after 2026-09-15.
What we found
- Out-of-sample direction AUC was 0.44 and 0.56, and about 0.51 in-sample: no directional skill.
- It detects volatility only and was no better than ATR14 (AUC 0.60 versus 0.57).
- With the veto on, the stop-out rate was 48.6% versus 55.1% for passed trades: no benefit. It was switched to record-only on 2026-10-17.
Caveats
- A new model is not used for decisions until it shows out-of-sample directional skill.
Volatility-based position sizing and the SMA200 filter
Limited or inconclusive · Tested
Idea tested: Scaling position size to volatility, or staying out below the 200-day average, improves risk-adjusted return.
Sample: 6 coins on daily candles since 2019, long only, 0.1% cost.
What we found
- None passed the pre-set criteria.
- Volatility-based sizing cut maximum drawdown by only 1 to 2 points, and its Sharpe ratio was below buy-and-hold (pooled difference -0.17, interval -0.38 to +0.02).
- Only the SMA200 trend filter cut drawdown substantially (-78% to -64%), but it reduced exposure to about 55% and returns by about half. The Sharpe difference was +0.04 with an interval including zero, and it failed on XRP and DOGE.
Caveats
- The 6 coins are survivors, which flatters buy-and-hold.
How to read this ledger
- These are results on past data. They do not predict the future, and a strategy that failed here may work in other conditions, and the reverse.
- Trying many ideas and keeping the ones that look good produces false positives. That is why we correct for multiple comparisons and still treat a single pass cautiously.
- A “promising” verdict is not a recommendation. It means a strategy is worth live paper testing, nothing more.
- Nothing here is investment advice.
Frequently asked questions
Why publish results that say strategies do not work?
Because most ideas that are widely shared do not survive costs and proper testing, and users deserve to know that before putting money behind them. AETHER’s AI assistant is built to quote this ledger instead of repeating popular claims.
Does any strategy on this page make money?
One family of hourly trend-following breakout rules was promising on historical BTC and ETH data, but it failed the drawdown criterion, is sensitive to costs and has not been tested live. Nothing here is a promise of profit.
What does “no edge” mean?
After costs, the strategy did not do better than chance in our test, or did worse. It does not prove that the idea can never work, especially where the sample was small.
What is R?
R is a unit of risk. One R is the distance between entry and stop-loss. An expectancy of +0.1R means an average gain of 10% of the amount risked per trade, after costs.
How often is the ledger updated?
When a new test is finished or an older one is corrected. Each entry shows the date it was tested.