Almost every systematic trader meets the same event: a strategy that earned steadily in the backtest bleeds in its first live weeks. The causes are knowable and rankable. Here they are, in the order that actually explains most cases.
The leading cause, and the least discussed by anyone selling something. A strategy tuned until history looked good has memorized noise, and live trading is the first data it could not memorize. The tell: live performance is not merely worse, it is characterless, hovering around zero minus costs. The fix happens before going live, not after: parameter-stability checks, an honest holdout, and selection-corrected statistics, covered in full in the overfitting guide.
A backtest that assumed fixed spreads, ignored slippage, or skipped funding charges overstates every marginal edge. Live, the strategy pays the real spread at the moment of each trade, slips on fast markets, and pays to hold. High-frequency and news-adjacent strategies suffer most, because their edges are smallest per trade. The test: re-run the backtest with costs doubled; if the edge vanishes, the live result was predictable.
Research data and your broker's feed are two different observations of the market. FX and CFD instruments have no single tape: each venue has its own liquidity, spreads, and timestamps, so a signal computed on one source can fire when the broker's price never touched the trigger, and fills happen at prices the research data never printed. This divergence is structural, not a defect, and Sonar Sciences treats it explicitly: research runs on consistent cross-venue data, execution fills at the broker, and the difference is stated rather than hidden. A robust strategy must survive small price perturbations, which is testable before going live.
Three losing weeks feel like proof of failure; statistically they are usually nothing. If the backtest's own distribution contains months like the one just lived, the live result is inside expectations and the correct action is patience by the numbers, not intervention. This is where Monte Carlo percentiles earn their keep: they define, in advance, what normal bad looks like, so live drawdown can be compared against a threshold chosen calmly.
Orders rejected without alarms, partial fills treated as full, a feed stalling and the system trading on stale prices, a retry sending the same order twice. Operational failures masquerade as strategy failures, and their signature is divergence between what the system believes happened and what the broker statement says. This is an infrastructure problem, and the reason Sonar Sciences's execution layer is built around typed, idempotent, auditable instructions: a failure must look like a failure, never like a losing trade.
When live diverges: reconcile broker statements against system logs first, execution bugs are the fastest to find and fix. Then compare live fills against backtest assumptions for the same period, which isolates cost understatement. Then compare live signals against signals recomputed on the broker's own prices, which isolates data divergence. Only after those three are clean is the strategy itself on trial, and the verdict should wait until the live sample is large enough to mean something.
In rough order of likelihood: the backtest was overfit to its history, trading costs were understated, your broker's prices differ from the research data, the live sample is still too small to judge, or execution has a silent fault. Diagnose in reverse: reconcile execution first, then costs, then data, and only then question the edge.
Some degradation is normal even for honest strategies: research always carries selection pressure, and live costs are real. Practitioners commonly plan for live performance meaningfully below the backtest and treat anything better as a bonus. A strategy whose economics only work at 100 percent of backtest performance is not robust enough to trade.
Yes, as filters rather than promises. An honest backtest with costs, adequate sample, stable parameters, and an out-of-sample pass dramatically raises the odds an edge is real. What no backtest does is guarantee the future; its job is to reject the strategies that would certainly have failed, which is most of them.
Bring one strategy you already trust. The Studio validates it against four years of real data, costs included, for free.
Start building