A backtest is a hypothesis, not a result. Here is exactly how Mercurio is validated — the universe, the execution model, the regime filter, walk-forward testing, and the drawdowns we refuse to hide.
Anyone can produce a beautiful equity curve. Overfit a few parameters to the last two years, ignore slippage, model stop-losses as if they always fill at the exact price, and you can manufacture a chart that climbs to the moon. That chart is a claim, not a result. The entire purpose of Mercurio's backtesting discipline is to tell the difference.
This article walks through how we validate the strategy that Mercurio actually runs. Every number here comes from the same engine — you can read it, and run it, yourself. The full backtest engine lives in portfolio_backtest.py on GitHub.
Mercurio runs a single edge: trend following on 1-hour bars. Entries require stacked exponential moving averages (12/48, 24/72, 48/144) confirming a direction, with an ADX floor confirming a trend actually exists. Exits use a Chandelier trailing stop (a multiple of Average True Range below the recent high) plus a catastrophe stop. Risk is capped at 1.5% of capital per trade. That is the whole strategy — deliberately simple, because complexity is where overfitting hides.
We test it on a universe of large, liquid US equities — names like MU, CRWD, PANW, ASML, and ARM — using $25,000 of paper capital, the same figure the live paper account is capped at.
This is where most backtests quietly lie. Two assumptions make or break realism, and we model both pessimistically.
Slippage. Every entry and exit pays 0.15% per side — roughly 0.3% round-trip — a more honest figure for liquid names than the 0.1% many backtests assume.
Gap-through-stop fills. A stop-loss does not guarantee you exit at the stop price. When a stock gaps down overnight, you fill at the open, not the stop. Modeling this — filling at the open price across session breaks rather than at the stop — was the single most important realism fix we made. It costs roughly 32 percentage points of return over our test window compared to the fantasy version where stops always fill exactly. We keep it on.
The difference between a backtest that assumes perfect stop fills and one that models gaps is the difference between a strategy you can trust and one that will surprise you with real losses. The gap model alone turned several 'winning' configurations into losers.
We did not take that on faith either — we attributed the cost knob by knob, holding everything else fixed over the same window. Slippage barely moved the needle; the gap model was the entire story:
| Execution assumption | Same window result | Effect |
|---|---|---|
| Stops always fill exactly (fantasy) | +26.6% | baseline (too optimistic) |
| Slippage 0.10% -> 0.15% per side | ~unchanged | noise, path-dependent |
| Gaps modeled on session breaks (live model) | session-gap | realistic exit fills |
| Gaps modeled on every bar (pessimistic) | -7.8% | over-penalizes intraday |
These are not hand-waves — they are environment variables in the engine (BT_GAP_THROUGH, BT_SLIPPAGE_PCT), so anyone can flip them and watch the number move. The honest configuration sits between the fantasy and the worst case: model gaps only across genuine overnight and weekend breaks, where they actually happen.
Mercurio only takes long positions when the broad market is in a confirmed uptrend: the S&P 500 must close above both its 50-day and 200-day moving averages for ten consecutive sessions. In bear or choppy regimes, the engine sits in cash.
While validating this, we found a bug that should make you suspicious of every backtest you have ever seen. Our engine had been computing the S&P's 50/200 'day' averages from 1-hour bars — roughly 7 and 28 trading days, five to seven times too fast. In a two-year bull market this barely shows; over a full cycle it is fatal, because the filter switches regimes far too eagerly. Fixing it to use genuine daily 50/200-day averages (mirroring the live bot) erased about 31 percentage points of phantom return from one multi-year test. We would rather find that ourselves than have the market find it for us.
Mechanically, the gate is a slow, deliberately conservative switch. The index has to close above both averages for ten straight sessions before a single long is allowed; anything less keeps the engine in cash. The backtest mirrors this exact logic so the simulated and live bots make the same decision on the same day.
The gate reads the S&P 500's own daily trend. Longs are only unlocked after the index closes above both its 50- and 200-day averages for ten consecutive sessions. Anything less means cash.
The cardinal sin of strategy development is tuning parameters until the backtest looks great, then being shocked when live results don't match. The defense is walk-forward analysis: optimize on an in-sample (IS) period, then test — untouched — on a later out-of-sample (OOS) period the optimizer never saw.
We ran exactly this. The in-sample period (2021–2024, including the 2022 bear market) and the out-of-sample period (2024–2026) were kept strictly separate. Most of the 'improvements' we tried — wider trailing stops, different exit widths, exposure caps — looked spectacular on one window and fell apart on the other. They were mirages. Only one structural change generalized across both windows:
That validated configuration — long-only, with the 10-day bull confirmation — is what Mercurio runs. The in-sample Sharpe clears the project's 1.0 bar; the out-of-sample Sharpe of 0.58 honestly does not, which is one reason we are still in paper validation rather than live.
To see why walk-forward matters, look at the tweaks that looked like winners on the full sample and what happened when we split the windows. A wider ATR stop and a faster Chandelier trail both posted huge full-sample returns — and both lost money in-sample, meaning their entire result was a recent-history artifact:
| Variant | Full-sample (naive) | In-sample | Out-of-sample |
|---|---|---|---|
| Validated: bull-only + confirm-10 | robust | Sharpe 1.13 | Sharpe 0.58 |
| Wider ATR stop (atr70) | +142% | Sharpe -0.07 (loses) | Sharpe 1.31 |
| Faster Chandelier (chand40) | +211% | -6.3% (loses) | +41% |
| Directional (long + short) | — | +9.9% | -6.9% |
Those +142% and +211% headline numbers are exactly the kind of figure a dishonest backtest would put on a landing page. Split the data and they collapse: a strategy that loses across 2021–2024 cannot be trusted just because it happened to fit 2026. Only the boring, consistent configuration survived both windows — and 'survives both windows' is the whole bar.
Replaying the live configuration over June 2024 to June 2026 with all of the realism above produces this equity curve. It is not a straight line — and that is the point.
A single return number tells you almost nothing about whether an edge is real or lucky. The honest way to read a strategy is through its trade distribution — and ours is unspectacular in exactly the way a durable edge should be. Across 381 trades, the average winner was $228.76 and the average loser was $210.99. The winners are barely larger than the losers; there is no single magic trade carrying the whole result.
From those four numbers the entire edge is recoverable by hand. Expectancy per trade is the win rate times the average win, minus the loss rate times the average loss:
(0.559 × $228.76) − (0.441 × $210.99) ≈ +$34.9 per trade
Multiply that $34.9 edge by 381 trades and you get about +$13,300 — which is, to the dollar, the difference between the $25,000 starting capital and the $38,278 ending balance. The profit factor of 1.37 says the same thing a different way: for every dollar the strategy lost, it made $1.37. Modest, positive, and consistent across hundreds of trades. That is what an edge actually looks like — not a hockey stick.
Notice the worst trade (-$3,203) is more than double the best (+$1,437). That asymmetry is the cost of gap risk: a stop can be jumped overnight, but a winner is trailed out in an orderly fashion. We model that pain rather than hide it — it is exactly why the gap-through-stop assumption matters.
Annual figures launder volatility. The monthly view is honest: the strategy lost more than 15% in its very first month, went flat for an entire quarter in cash, and posted its single best month (+21%) only at the very end. A real account lives through every one of these bars in sequence.
Trend following has fat-tailed returns by construction, and the per-symbol breakdown proves it. Five names — MU, CRWD, PANW, ASML, and ARM — produced about $10,040 of profit between them, roughly three-quarters of the entire net result. A long tail of small winners and losers fills in the rest. This is a feature, not a flaw: the strategy exists to stay in the few trades that become trends.
| Top contributors | Net P&L | Worst contributors | Net P&L |
|---|---|---|---|
| MU | +$2,470 | ZS | -$964 |
| CRWD | +$2,195 | PYPL | -$836 |
| PANW | +$1,956 | XOM | -$687 |
| ASML | +$1,737 | GE | -$671 |
| ARM | +$1,680 | — | — |
It also flags the central risk: concentration. If those five names had trended sideways instead, the two-year result would be roughly flat. A handful of winners carrying the portfolio is the same property that produces the deep drawdowns — you cannot have one without the other.
Notice that the curve falls before it rises. Through late 2024 the portfolio sank to about $18,175 — roughly 27% below its starting capital — and spent months underwater before recovering. The maximum peak-to-trough drawdown over the window was 29.3%. A real person watching $25,000 fall below $18,200 has to keep their nerve. Any backtest that shows you only the smooth final number is hiding the part that actually tests you.
We tried hard to engineer that drawdown away — exposure caps, a close-on-breach circuit breaker, volatility targeting. Each one either failed to help or made the out-of-sample result worse. The drawdown is a structural property of long-only momentum on a tech-heavy universe during sharp corrections. We document it rather than pretend we solved it.
Over a full five-year cycle (2021–2026) that includes the 2022 bear market, the same strategy was net negative: about -23% total, a negative Sharpe, and a 52% drawdown. Trend following on this universe works in sustained bull markets and struggles everywhere else. The strong two-year window is real, but it is not the whole story.
This is why the regime filter exists, why we are long-only, and why we sit in cash for long stretches. It is also why Mercurio is in a 12-month paper trial instead of trading live capital. We are testing whether the validated configuration holds up going forward — not assuming it will.
None of this asks for your trust on faith. The engine, the strategy, the risk manager, and the regime logic are all on GitHub. The backtest is deterministic: same configuration, same window, same numbers.
backend/engine/portfolio_backtest.pyBT_GAP_THROUGH, BT_SLIPPAGE_PCT, BT_REGIME_SMA, BT_BULL_CONFIRM_DAYS.backend/config.py and are enforced by backend/engine/risk_manager.py.If you change one realism assumption — turn off the gap model, speed up the regime average — you will see the numbers improve, and you will understand exactly why we don't.
Disclaimer. Every figure in this article is from a historical simulation using paper capital. It is not live trading, not a forecast, and not financial advice. Past performance does not guarantee future results. Trading involves substantial risk of loss.