The goal of this project is an automated trading system whose logic and results are public. To build one you have to know what you are choosing from. This is the map we work from: every family of FX technique we know of, and what our own backtests say about it. Nothing here is a claim about the literature. Every status is a run you can open on the presets page, with its parameters and its trade count.
A word on how to read "status". We tune on 2025 to September 2026, then run the same values on every earlier window we have, back to 2019. A technique that only works on the tuning window is not a technique, it is a fit. A technique that works on 2023–2026 but loses on 2019–2022 is what we call a regime edge: it may be run at minimum lot with a kill switch, but it is not proven. Only something that holds on every window is a mechanism edge. As of today we have several regime edges and no mechanism edge.
The map
| Family | What it bets on | Status in our tests |
|---|---|---|
| Mean reversion (bands, RSI, VWAP) | Price snaps back to a local average | The only edge found. Buying the lower-band dip in the European morning on M15, long only, three to four pairs. Profit factor 1.3–1.7 on 2023–2026, below 1.0 on 2019–2022: a regime edge, live at minimum lot. |
| Trend following (moving-average cross, ADX, slope) | Moves persist | Rejected. Profit factor 0.53–0.93 across pairs on M15 and H1; a 25-pass exit sweep was negative on every pass. |
| Pullback in a trend | Re-enter after a dip inside a trend | Rejected on two pairs. Same failure as trend following. |
| Range breakout (session range, opening range, Donchian) | A break out of a range continues | Rejected intraday: 954 trades, profit factor 0.61–0.88, both sides. The London/New York overlap chops; breaks neither run nor revert cleanly. |
| Time-series momentum on daily or weekly bars | Multi-week persistence, holds of days to weeks | Not tested. Few trades, but we have daily bars from 2013. |
| Fading extremes (RSI extremes, daily overextension) | Overextension reverts | Candidate, thin. Fading our strongest daily Breakout reading gave profit factor 1.1–1.14 and about nine pips a trade, flat before 2020. Not traded. |
| Session and time of day | Liquidity and participants differ by hour | The European morning (07–13 server time) is where the reversion works. The same rules in the Tokyo morning and the New York session lost. |
| Fix-time flows (Tokyo 9:55 JST fix, London 16:00 fix, month-end fix) | Known flows at known times | Not tested. Well documented, and the cause is a real participant, not a pattern. |
| Calendar effects (turn of month, day of week, year-end yen flows) | Recurring flow patterns | Not tested. Cheap to test with a large sample. |
| Carry and rate differentials | Earn the swap, trade with the rate gap | Not tested. The tester applies today's swap to history, so this needs an external rate series. |
| News and events (breakout, fade, blackout) | Scheduled volatility | Not tested as an entry. The events calendar exists so that the engines can eventually use it as a filter. |
| Volatility regimes (ATR, ADX, vol targeting) | Filter or size by volatility | Used as a filter, and it matters: an ADX ceiling of 34 and an ATR floor are part of why the reversion entry works at all. Above 34 the same entry admits losers. |
| Liquidity sweeps and stop runs, round numbers | Stops cluster, get run, price reverses | Not tested directly. The morning dip we trade may be exactly this, which would make it worth generalising. |
| Price-action patterns (pin bar, engulfing, inside bar, support and resistance) | Candle shapes carry information | Not tested. Easy to encode, usually weak on their own. |
| Multi-timeframe confluence (daily bias, intraday entry) | The higher timeframe sets the side | Rejected both ways. Our daily readings are contrarian, and gating the intraday entry on them, with or against, removed trades without adding edge. |
| Correlation and baskets (yen strength, dollar index) | Trade the strongest against the weakest | Not tested. |
| Statistical and machine-learning regime models | Classify the regime, switch strategy | Not tested. Our sample sizes are small for it, and a regime model that cannot be read is hard to publish honestly. |
| Scalping | Very short holds, spread capture | Not viable on our data. Real tick history starts in 2026; M1 tests of the reversion entry lost (profit factor 0.59), and the spread is most of the move. |
| Grid, martingale, hedging schemes | Not an edge, a payout shape | Deliberately not tested. They turn many small losses into a rare large one, and a track record cannot show that until it happens. |
| Exit techniques (trailing, partial, time stop, mid-band exit) | Improve an existing entry | Tested on the reversion entry: the mid-band exit and a four-hour time stop are in use; trailing stops changed nothing. |
Two things the tested column says
Everything that assumes continuation failed at intraday horizons. Trend following, pullbacks, breakouts, daily-bias gating: every one of them lost on the majors on the timeframes we can test, in both sessions we tried. This is consistent with the forward-return study of our own daily signals, which found that a strong directional reading is a measurement of how far a move has already gone, not a forecast that it continues.
Everything that made money assumed reversion at a known time and place. A specific session, a specific band, a volatility ceiling. The narrower and better-explained the setup, the better it held out of sample. That points at where a mechanism edge might live: not in a pattern but in a cause, a participant who has to trade at a known time, or a stop cluster that has to be run.
What we test next
In order of how plausible the cause is:
- Fix-time flows. The Tokyo fix on USDJPY, the London 4pm fix on the European crosses, and month-end fixes. Real participants on a real clock, testable on our data today.
- Liquidity-sweep reversal. A push through the previous session's low that closes back inside the range, then buy. If this is what the morning dip really is, the rule should work in more places than one session.
- Turn of month and day of week. Large sample, cheap to run, and either the effect is there or it is not.
- Daily and weekly time-series momentum. The one continuation idea at a horizon where it is known to exist, on thirteen years of daily bars.
Each will be filed on the presets page the same way as everything above: the values, the windows, the trade counts, and the verdict, whether it passes or not. Most will not. That is the point of keeping the map.
Nothing on this site is investment advice. Every result above is a backtest at 0.01 lots on one broker's data, published so that it can be checked.
评论