Research log · crypto and forex · 2026
23,600 strategy tests. After costs, almost nothing is left.
One person's systematic search for a trading edge, on real market data, with realistic costs and, where it came to live trading, on demo accounts only. It covers a foundation-model forecaster, 162 technical indicators, 31 on-chain and macro predictors, carry, intraday patterns and scalping. Most of it failed, and the failures are the useful part.
≈23,600configurations tested
0significant out-of-sample edges after costs
2bots running on demo accounts
What we would tell a friend
- Costs decide. The edges we measured were usually 5 to 50 times smaller than a round trip. Start with the cost hurdle.
- In-sample winners fade. Every strategy that entered our sealed 2025–26 hold-out lost money, including a trend portfolio with a training Sharpe ratio of 2.0.
- A high win rate is not an edge. Rules that won 60–80% of their trades lost money after costs.
- What little works is momentum and risk control. A direction model and a trend vote halved drawdowns in crypto, but neither beat simply holding on return with any statistical confidence.
- Retail currency trading through CFDs starts at a loss. The swap mark-up alone, 1–1.8% a year per side, is about the size of the published currency premia.
Scoreboard
Every strategy family we tested, with its best out-of-sample evidence after costs. Click a name for the full write-up.
| Strategy family | Market | Variants | Out-of-sample, after costs | Verdict |
|---|---|---|---|---|
| Kronos foundation-model forecasts | BTC, ETH, SOL | 43 | Skill only at 5–15 minutes (IC +0.086); 0 of 43 cells profitable after fees | Failed |
| Classic rules and coin portfolios | Crypto, FX | 4,032 | All six hold-out finalists lost; the best −35.1% against −16.3% for holding Bitcoin | Failed |
| Buy the dump, fade the pump | 12 coins, 1-minute | 3,456 | 80% winning trades, −6% to −11% in the hold-out | Failed |
| Candlestick and chart patterns | Crypto, FX | 2,814 | 0 tradeable; the two confirmed patterns worked backwards | Failed |
| Calendar, weather and the Moon | Bitcoin, daily | 40 | 0 of 40 survived confirmation | Failed |
| 162 technical indicators | 20 instruments | 9,720 | 0 of 162 after correction; always long beat every indicator | Failed |
| Crypto trend vote | 11 coins, 4-hour | 7 | Sharpe 0.51 against 0.16, drawdown −39% against −75%; alpha t 0.8 | Risk filter |
| Intraday and cross-pair FX | 9 instruments, 1-minute | 1,442 | 0 passed; gross edge about zero | Failed |
| FX carry | 16 currencies | 8 | 5.0% a year before swap mark-ups, 2.1% after (t 1.0), −56% drawdown | Weak |
| 2-year yield-gap momentum | 7 currencies | 4 | Sharpe 0.25 before costs and 0.02 after in 2014–26 | Weak |
| FX positioning, trend, FOMC, month-end | G10 currencies | 40 | All negative after costs in the recent period | Failed |
| EUR/USD scalping | EUR/USD ticks | 1,630 | 0 profitable in both halves; live −19.62 USD over 201 trades | Failed |
| Order-book imbalance | EUR/USD ticks | 11 | Real (t 16) but 0.01–0.04 pip against a 0.9-pip cost | Failed |
| Broker depth ladder skew | EUR/USD depth | 1 | Pre-registered; first run on 3 October 2026 | Pending |
| Stablecoin inflows | Bitcoin | 1 signal, ~40 exits | +1.6 to +2.3 pp per signal since 2021 (t ≈ 2); did not beat holding over 2019–26 | Candidate |
| Direction model, 5 families | BTC, ETH, 9 altcoins | 70 | Sharpe 1.22 against 1.15, drawdown −34% against −53%; alpha not significant | Risk filter |
Posts
- The cost hurdleWhy almost every strategy on this site fails. The edges we found are measured in hundredths of a percent per trade, and a round trip costs tenths.
- Can a foundation model forecast crypto?We ran the open-source Kronos candlestick model through a full year of Bitcoin, Ether and Solana. It has real skill at 5 to 15 minutes, and that skill is about twenty times too small to pay the fees.
- Classic rules, one vault and 14,000 backtestsMoving averages, breakouts, dip buying, candlestick patterns, the Moon and macro filters. Everything that looked good on 2020–2024 data lost money in a sealed 2025–26 hold-out.
- 162 technical indicators against 20 random rulesWe backtested 162 textbook indicators and candlestick patterns on 20 markets and three timeframes. None beat random rules after correcting for multiple testing, and simply staying long beat all of them.
- Forex on a retail CFD account: 3,100 tests, no edge after costsIntraday patterns, cross-pair lead-lag, carry, positioning data, FOMC days and yield momentum. Several of these are documented premia. On a retail CFD account the commission and the swap mark-up remove every one of them.
- Scalping EUR/USD live: fast fills, no edgeWe ran four scalping strategies on two broker demo accounts and studied the order book. Orders filled in under 300 ms with no slippage. The strategies still lost, because a 0.7-pip round trip is bigger than almost anything they could predict.
- Stablecoin inflows and Bitcoin: a modest, recent effectWhen total stablecoin supply jumps, Bitcoin tends to do slightly better over the following week. Since 2021 the effect is 1.6 to 2.3 percentage points and shows up in three data sources. Before 2021 it was a data artefact, and the simple rule did not beat holding Bitcoin.
- Thirty-one ways to predict crypto directionWe tested 31 published predictors of Bitcoin and Ether direction. Most of what works is momentum in disguise, the contrarian indicators point the wrong way, and a simple five-family average beats holding only on risk.
- Live results on demo accountsTwo bots run around the clock on demo accounts: a daily Bitcoin bot driven by stablecoin inflows and a EUR/USD scalping bot. This page is a snapshot of what they have done so far.
About this log
The research ran in September 2026 on public data and broker demo feeds. Every number on the site comes from the project’s own code and result files, and every chart has a data table underneath. The studies, the tests and the text were produced by one person working with an AI coding assistant. The methods page lists the rules we held ourselves to, and the data page lists every source.
Nothing here is investment advice. No real money was traded.