the whole table
every strategy, not the winner
| strategy | in sample | out of sample | equity, out of sample | ||||
|---|---|---|---|---|---|---|---|
| return | max dd | sharpe | return | max dd | sharpe | ||
| rsi(14) mean reversion | -39.7% | -70.9% | -0.08 | -24.2% | -55.0% | -0.15 | |
| 50/200 moving average | -62.6% | -82.7% | -0.42 | -15.1% | -36.6% | 0.02 | |
| 20 day breakout | -77.3% | -77.3% | -0.51 | 39.4% | -39.2% | 0.76 | |
| 12 month momentum | -61.2% | -72.7% | -0.82 | -29.5% | -29.8% | -0.99 | |
| buy and hold | -67.9% | -88.3% | 0.00 | -61.5% | -84.1% | -0.36 | |
the in sample winner was buy and hold at -67.9%. picking it on that evidence then cost 61.5% in the window that followed. that is the finding, and it is why this page prints all five.
the bot verify on chain
what ours does instead
the table above measures returns on one asset. the bot is measured differently, so it gets its own block rather than a row: it runs one rule on a funded account, and what it is judged on is whether the account clears, and how long that takes. putting a pass rate in a returns column would be comparing two different things.
| pace | passes | stopped out | average | 9 in 10 take |
|---|---|---|---|---|
 slow | 45% | 55% | 50 days | 10 to 105 days |
 medium | 35% | 65% | 25 days | 5 to 70 days |
 fast | 30% | 70% | 15 days | 5 to 35 days |
the pace is the risk per trade and the only thing that changes: the strategy under all three is the same one. a faster pace passes sooner and less often, and the stopped out column sits beside the pass rate because we are the counterparty and the bot supplier at once.
reading it
what this says
arbitrum fell 61.5% through the out of sample window, and exactly one rule came out ahead: 20 day breakout, at 39.4%. the honest reading is not that these five are bad, it is that a backtest chosen on past performance tells you little about the next window.
the one rule that came out ahead, 20 day breakout, was in the market 37% of the window. it earned its 39.4% by being out of it for the rest, which is a timing claim that one window cannot settle.
method
how it was run
- data
- arbitrum daily candles, hyperliquid public api, 1,315 bars, 2023-03-23 to 2026-10-27, read 2026-10-26.
- fees
- 5 basis points a side on every change of position. funding is not modelled, and a long pays funding most of the time, so real figures would be slightly worse, never better.
- no hindsight
- every decision is taken on one close and applied to the next day's move.
- the split
- first 70% sets expectations, last 30% tests them. nothing was tuned on it.
- missing
- one asset, one timeframe, long only, no sizing, no stops.
questions
what people ask
what is the best strategy for arbitrum?
20 day breakout came out best at 39.4%, against buy and hold at -61.5%. 1 of the five made money out of sample.
do these numbers include fees?
yes, 5 basis points a side on every change of position. funding is not modelled.
why show strategies that lost?
because showing only the winner is how backtests lie. fifty strategies across a hundred assets is five thousand results, and the best of those is mostly luck.
if this looks like work
it is, and we did it
picking a rule, testing it honestly, sizing it, then holding through a 84.1% drawdown is most of a job. ours runs one rule at one pace on a funded account, and we print how often it passes beside how often it stops out.
more