Since September we have written up 28 studies, most of them looking for something new to trade. Exactly one mechanism has earned our proven tier: the Tokyo fix. This is a report on the rest, what they had in common, and what we now do with strategies that work for a while but can't be proven.
What we tested
You will have seen most of these ideas sold somewhere. They failed in a few repeating ways.
| Family | Examples | What happened |
|---|---|---|
| Chart patterns and indicators | Breakouts, liquidity sweeps, RSI and MACD gates, our own daily signals as an input | Nothing above the spread that held in both halves of the data |
| Trend following | Time-series momentum, multi-asset trend, trailing exits | Looked fine in the years we built it on, then lost in the years we held back. Trailing stops made it worse. |
| Currency relationships | Lead-lag between pairs, currency strength, pair spreads | Mostly nothing. EUR/GBP looked good from 2010 and did nothing on 2003-2009. |
| Calendar effects | Turn of the month, start of the month, weekend gaps | Too small after costs, or the sign flipped |
| Scheduled flows | London 4pm fix, ECB fix, PBoC fix, Tokyo and New York option cuts, month-end hedging | A clock is not enough. Most were two-sided or already anticipated. |
| Published anomalies | The overnight drift in US indices | Fell to zero after 2022 on our broker's data, matching what the New York Fed now reports |
Our two most recent studies make the point. For the New York option cut, we traded towards the nearest round number before 10:00 ET, on the theory that big option strikes pull the price in. Before costs it made −0.10 pips a day, the same as a control half-hour with no cut. After the spread, every pair and every year lost. For the overnight drift, the famous hour around the European open earned +1.5 basis points a day in 2020-22 and +0.05 since.
Why the survivor survived
The Tokyo fix has three things the others lacked: a clock, a participant who has to trade, and a known direction. The usual story is that Japanese importers have to buy dollars at the 09:55 fix on set dates, whatever they think of the price. We can't see those orders. What we can measure is the fall in the yen pairs afterwards, in every full year since 2019.
Almost everything else was missing at least one of the three. Option expiry has a clock and a participant, but the flow runs both ways. A moving average crossover has none of them.
Why adding conditions rarely helps
When a strategy fails, the tempting fix is another filter. It nearly always looks better, because you can see which trades lost and pick a filter that removes them. We tried this on the Tokyo fix itself: no condition we could observe at 09:55 isolated its losing days.
Scanning has the same trap. The overnight study printed all 22 trading hours for the record, and the best was 11:00 to 12:00 ET. We first read it as a t-statistic of 3. That counted two closely linked indices as separate evidence. Counted properly it is 1.9, and it is the best of 22 tries, which is what luck looks like. We froze the rule anyway and will judge it only on days the scan never saw. That needs about 1,000 trading days.
A new use for an old label: regime
Some strategies sit in between. They have worked in recent years and clearly did not before. Our London RangeRevert EAs are like this, and so is a month-end GBP/USD fade we rejected in September: profit factor 1.50 for 2019-26, 0.88 for 2010-18.
Calling those proven would be false. Throwing them away ignores what they are. So they run under the regime tier, which means working for now, not proven, at minimum lot.
The obvious plan is to switch them off when the wind changes. We tested that too, and it failed: rules that turn a strategy on and off from its own recent results did not pass on any measure. By the time results look bad, much of the damage is done.
So we don't pretend to time it. The GBP/USD fade got stop rules written before its first live trade:
- A loss budget. JPY 3,000. Reach it and the run ends.
- A losing-streak limit. Seven in a row.
- An expiry. 24 live trades or November 2028, whichever comes first. Doing nothing means it stops.
- A reason check. If the fix itself changes, the run ends even if it is winning.
Those rules are on the EA's page. They don't predict anything. They set in advance what the experiment may cost. In September we said this one wasn't worth running at twelve trades a year, and for income that is still true: at 0.01 lot it has made about 87 pips since 2023 in backtest. What it can earn is a live record.
Where things stand
- Proven and live: the Tokyo fix on four yen pairs.
- Regime and live: four London RangeRevert EAs and the GBP/USD month-end fade.
- Forward tests: the US-close fade, pre-FOMC drift, EUR/GBP rubber band and the 11:00 ET hour. All are too young to mean anything, and none is being traded.
- Live record so far: 23 closed trades on the current live EAs, which is far too few to judge, whatever the number says. It is on the track record page.
What this does not tell you
It doesn't say these markets have no edges. It says 28 studies on retail data, at retail costs, found one we can stand behind. Some of what we closed might work with better data, such as real option-strike information. The Tokyo fix could also fade: proven describes the past. And a backtest is one broker's history, not a promise.
We will keep looking, but less often, and only when an idea has a reason behind it. The results, failures included, stay on the presets page.
Nothing here is investment advice. Trading carries a real risk of loss.
Komentar