Why Trading Systems Stop Working: What to Know About Market Regimes
Markets change character in phases, so a system works some years and breaks in others. Evidence from gold M1 2022–2026, and how to read multi-year tests honestly.
- Markets go through phases (regimes). A system that works in one phase can break in another, even though its rules aren't wrong.
- A good-looking multi-year total may come from a single year. Always split results by year.
- A threshold set in points or dollars changes meaning when price changes. Use % of price instead.
- Trend-following and counter-trend systems tend to work in different years.
- Write your pause rules in advance, and cut trade size when you're unsure the market is still the same.
When a trading system that used to work stops working, it’s usually not because the rules were wrong, but because the market changed character. We call these phases of market behaviour a Market regime, for example a phase where price runs in long trends versus a phase where it swings inside a range. One system usually fits one regime. This article looks at the evidence from our tests on gold, then sums up how to read results across years and how to cope.
What a regime is, and why systems “expire”
Every system has a hidden assumption. A trend-following system assumes price that breaks out will run far. A counter-trend (or fade) system assumes price that overshoots will come back. If the market really runs far in that phase, the first system profits and the second keeps getting stopped out. When the market goes back to swinging, the results swap.
Every line of the rules stays the same. What changes is the market.
Evidence from gold: the wall between 2022–24 and 2025–26
In our backtests on gold (XAUUSD) M1, we saw the same picture again and again. Of the systems that survived to the final round, whether trend-following, pullback or range breakout, almost all lost money or just broke even in 2022–2024, then made a profit in 2025–2026. Systems built on different ideas broke in the same years and recovered in the same years. That means the deciding factor was the market, not the formula.
One example: a pullback system that uses turning points (fractals). We tried 7 parameter values. In 2026, all 7 were positive. Across 2022–2025, 25 of the 28 cells lost money (1-minute data). No amount of tuning helped, because the real variable was the year.
Systems that pass across several years still lean on one year. Every one of our Fibonacci pullback systems that was positive across several years had 70–86% of its profit in 2026 alone.
One reason is how often price runs far enough. For a pullback system at 61.8% of large swings, counting legs where price ran $30: 2022 had 81, and 2026 (January to September) had 1,249, or 1,249 ÷ 81 = 15.4 times as many, even though 2026 counts only 9 months.
All results in this article are backtest results. Past results do not predict the future and are not a return you should expect.
Trend-following and fade systems work in different years
The clearest picture comes from putting two systems side by side. Both use the same ZigZag swings on M1.
- Trend-following system enters with the move when an M1 bar closes beyond the end of the previous swing. Wide Stop loss, far target.
- Morning fade system fades when price runs past 161.8% of the previous swing, only from 07:00–11:00 Thai time (UTC+7).
Conditions: gold M1, 0.10 lot per trade, $10,000 starting balance. 2024–2025 use 1-minute price data; 2026 uses real ticks from January to September.
In 2024 the market swung in ranges: the fade system made a profit, the trend-following system lost, and its win rate fell to just 9%. When gold trended hard in 2025–2026, the results flipped at once.
Think about how much the trend-following system depends on one year:
- 3-year total = −2,764 + 7,404 + 17,836 = $22,476
- 2026’s share = 17,836 ÷ 22,476 = 79%
- Check: remove 2026 and you’re left with −2,764 + 7,404 = $4,640, and 4,640 + 17,836 = 22,476, which matches the total
Someone who started the trend-following system at the start of 2024 would have lost for a whole year before seeing profit. Someone who started the fade system at the start of 2025 would have seen nothing but losses. See also Ride the trend or fade it.
A system isn’t simply “good” or “bad.” It’s good in one kind of market. Before you trust a total, ask which year the profit came from.
Point-based thresholds quietly change meaning
A regime doesn’t just change direction; it changes the size of moves too. We measured the median width of large ZigZag swings on gold M1 for each year.
| Year | Width (points) | Width (% of price) |
|---|---|---|
| 2022 | 13,564 | 0.76% |
| 2023 | 12,148 | 0.64% |
| 2025 | 28,412 | 0.89% |
| 2026 | 49,640 | 1.18% |
2024 isn’t in this measurement set, so we left it out.
In points, 2026 is 49,640 ÷ 13,564 = 3.7 times wider than 2022. But as a % of price, it’s only 1.18 ÷ 0.76 = 1.6 times wider. The rest of the difference comes from gold getting much more expensive.
The result: a threshold set in points changes meaning on its own without anyone editing it. In our measurement, the rule “swing must be at least 80,000 points wide” passed only 2% of swings in 2022 but 41% in 2026. The same line of rules became a different filter.
Example: a $20 threshold vs a 1% threshold
For gold in the account we tested, 1 point = $0.001. We can estimate the average price from the table:
- 2022: 13,564 points = $13.564, which is 0.76% of price, so price ≈ 13.564 ÷ 0.0076 ≈ $1,785
- 2026: 49,640 points = $49.640, which is 1.18% of price, so price ≈ 49.640 ÷ 0.0118 ≈ $4,207
Say the rule is “only enter swings at least $20 wide.”
- 2022: 20 ÷ 1,785 = 1.12% of price, stricter than a typical swing that year (0.76%)
- 2026: 20 ÷ 4,207 = 0.48% of price, much looser than a typical swing that year (1.18%)
Check: 1,785 × 1.12% ≈ 20.0 and 4,207 × 0.48% ≈ 20.2, both close to $20 (the gap is rounding).
Switch to “at least 1% of price” and the threshold moves with price by itself: about $17.85 in 2022 and $42.07 in 2026. In our measurement, the 1% threshold let swings through every year, about 193–295 swings per year. A % threshold still isn’t perfectly equal, because gold really did become more volatile, but it stays much closer to its original meaning.
Same system, different markets
Regimes don’t just differ between years; they differ between assets too. We ran the set of systems developed on gold on 11 other symbols, adjusting lot and spread for each, but not the rules, using real ticks from January to September 2026. We looked at Profit factor (total profit divided by total loss; above 1 means profit, and it doesn’t depend on lot size).
| Symbol | Profit factor |
|---|---|
| XAUUSD (gold) | 2.01 |
| GBPUSD · CHFJPY | 1.09 · 1.06 |
| USDJPY | 1.00 |
| XAGUSD (silver) · GBPJPY · EURUSD | 0.96 · 0.93 · 0.89 |
| EURJPY · AUDJPY | 0.82 · 0.82 |
| AUDCHF · EURAUD · GBPAUD | 0.69 · 0.66 · 0.63 |
Only gold had a clear edge. Another test on M15 showed the same thing: a pullback system on major currency pairs won 35–41% of the time, close to gold’s 38.8%, but gold’s Profit factor was 1.34 versus 0.91–1.01 for the currency pairs. Entries were about as good, but after entry gold ran further. So the edge belonged to the market’s character, not to the formula.
How to read multi-year tests, and cope when the regime changes
- Always split results by year. A multi-year total can hide the years that broke. Check what % of the profit came from one year. If it’s more than half, treat it as unknown what the system will do in a year when the market is different.
- Measure in % of price, not points. That goes for Stop loss distance, profit target and swing-size thresholds.
- Cut trade size when unsure. If a system only did well in the recent period, start with a smaller lot than the backtest tempts you to use. For how to work out the lot, see Position size from Stop loss distance.
- Write your pause rules in advance. For example: “If the running loss goes beyond the level seen in the backtest, stop and review.” Write it while you’re calm, not while you’re losing, and keep it separate from losing streaks that are normal.
- Don’t tune the system to the latest year. You’ll get values that fit only that one year. See Overfitting.
A good stop rule is one written before the event happens. If you write it while you’re losing, it’s just emotion.
Limitations: all of this was measured mainly on gold M1. 2022–2025 are 1-minute data; only 2026 is real ticks (Why tick data quality matters). Part of this wall is also about timeframe: one system, moved to M15, turned positive in 3 of the 4 years from 2022–2025. We don’t know how long a market like 2025–2026 will last.
Common mistakes
- Trusting a multi-year total. A system with 79% of its profit from one year will look good in the total until the market changes.
- Changing the system after every loss. If the market has changed, no amount of formula editing helps, and the more you edit, the more it fits the past.
- Setting point thresholds and forgetting them. An 80,000-point threshold passed 2% of swings one year and 41% another, without anyone touching the code.
- Moving a system to another asset without testing. What works on gold may lose money on currency pairs.
Frequently asked questions
Should I drop a system that has lost for several months in a row?
First check whether the loss goes beyond the range you saw in the backtest. If it's still within that range, it may be a normal losing streak. If it goes past the limit you wrote down in advance, stop as your rules say, not based on how you feel that day.
Is splitting results by year enough, or do I need monthly?
Yearly is the minimum, because it shows the market's phases clearly. Monthly helps you see whether profit is bunched into a few months, but the sample per month is usually too small to conclude anything.
Can I know in advance when the market will change regime?
We don't have a way to know in advance. What you can do is measure the current market, such as swing size as a % of price, and prepare rules for what you'll do if the numbers change.
This is general education, not investment advice. Trading is risky and you may lose all of your capital.