We built a dozen strategies and backtested them honestly. Eleven failed. This is the story of how Mercurio went from a diversification thesis to a single, focused edge.
Mercurio began with the most conventional idea in quantitative trading: diversification. Run many uncorrelated strategies — trend, mean reversion, momentum, gap, pairs, plus crypto variants — and let them smooth each other's returns. It is textbook portfolio theory. It is also, for a retail-scale bot on a single broker, mostly wrong.
We didn't reason our way to that conclusion. We built every one of those strategies, then put them through the same honest backtest — realistic costs, multiple regimes, out-of-sample testing. The results were brutal and clarifying.
Eight strategies, the same universe, the same realistic execution model, the same multi-regime windows. Here is how they actually scored — not in theory, in the backtest:
| Strategy | Measured win rate | Verdict |
|---|---|---|
| Trend following | ~56% | The only durable, positive edge |
| Mean reversion | ~49% | Profitable in calm tape, gave it all back fading real trends |
| Gap trading | 28.6% | Caught noise, not gaps |
| Pairs trading | 27.3% | Correlations broke exactly when it mattered |
| Momentum scalping | 0% | Died entirely to spread and slippage |
| Crypto (TF + MR variants) | net negative | Losers across every variant tested |
Every non-trend strategy failed to show a durable edge. Mean reversion was the most seductive failure: it looked profitable in calm markets, then handed it all back the moment a trend it was fading kept going. Momentum scalping had a literally 0% win rate once realistic costs were applied — the edge existed only in a frictionless fantasy. The crypto variants were net losers across the board.
Several strategies also carried stacked 'conviction' multipliers on their position size — a chain of small reductions like 0.8 x 0.9 x 0.7. That looks harmless until you multiply it out: 0.504. Half the intended size, gone, before a single trade. A handful of plausible-sounding adjustments had quietly throttled the book to nothing. Reductions multiply; they do not average.
Trend following wins for a structural reason: markets have fat tails. Most moves are noise, but a small number of moves are enormous, and they cluster into trends. A strategy that cuts losers quickly and lets winners run is positioned to capture those rare large moves. Its win rate doesn't need to be high — Mercurio's is around 56% — because a positive expectancy plus a stream of trades compounds in your favor.
The per-symbol breakdown of our validated two-year backtest shows the fat tail directly. Just five names produced roughly $10,000 of profit — about three-quarters of the entire net result — while everything else netted out small. The strategy's whole job is to still be holding those five when they run.
Crucially, trend following also fails gracefully. In a choppy market it bleeds small losses; it doesn't blow up. The yearly returns make the dependence on regime impossible to miss — a losing 2024, then two strongly positive years once the trend resumed:
Combined with a regime filter that keeps it in cash during bear and sideways markets, those small losses in the wrong years become 'no trade at all.' The strategy doesn't try to be right in every regime; it tries to be present in the regime where it works and absent everywhere else.
It would be easy to stop at the +53% two-year figure and call trend following a winner. That would be dishonest, because the two-year window is mostly a bull market — exactly the weather this strategy is built for. So we ran the raw strategy, without the cash-in-bear discipline, over a full five-year cycle that includes the 2022 bear. The result is brutal:
Broken down by year, the dependence on regime is stark. 2021 (bull) made about +$2,750. Then 2022 (bear) lost -$7,333 — and the worst of that was the strategy's short side: 74 short trades lost -$5,175 between them. Trying to flip trend following around and ride the market down simply does not work on this universe. 2023 (choppy) lost again; 2024 was roughly flat; only the renewed bull of 2025-26 recovered.
Trend following has a real edge in sustained bull markets and a negative one almost everywhere else. You cannot fix that by making the entries cleverer; the only fix is to not trade when the weather is wrong. That single conclusion — be present in bull, flat otherwise — is why Mercurio is long-only, gated by a 10-day bull confirmation, and still in a 12-month paper trial rather than risking real money.
Focus also means resisting changes that feel safer but cost money. We added a filter that pauses new long entries during sharp market dips — it makes the ride more comfortable. When we measured it honestly, that comfort cost roughly 16–18 percentage points of return over two years versus leaving it off. We kept it anyway, as a deliberate, documented trade-off for steadier behavior — but we know its price, because we measured it. That is the difference between a preference and a delusion.
No strategy gets added back, and no new strategy gets added, without a fresh backtest proving a positive, durable edge across 2+ years and multiple regimes. Disabled strategies stay in the codebase for history — they do not trade.
The result is a system that is easier to reason about, easier to risk-manage, and honest about where its edge comes from. One idea, executed with discipline, beat a portfolio of mediocre ones. You can read the full validation in our backtesting deep dive.
Disclaimer. Performance figures are historical simulations on paper capital, not live results or financial advice. Past performance does not guarantee future results.
