A trading strategy can produce an attractive backtest and still lose money when real orders reach the market. The gap may come from an overstated historical edge, unrealistic execution assumptions, changing conditions or decisions that differ from the tested rules. “Profitable” needs a qualifier: profitable on which data, after which costs, and under what conditions?
The useful response is not to replace the strategy after every losing week. It is to identify where the evidence stops matching reality. Start by separating three questions: did the strategy have an edge, can that edge survive trading costs, and are you actually trading the same system?
The Backtest May Have Found Noise, Not an Edge
Testing hundreds of indicator settings and selecting the best result creates a selection problem. Some combinations will look impressive through chance alone. The research paper The Probability of Backtest Overfitting examines how selecting strategies from repeated historical tests can produce disappointing performance on unseen data.
Suppose a breakout strategy works with a 17-period lookback but struggles at 16 and 18 periods. That narrow sweet spot deserves scrutiny. Ask what market behaviour the rule is meant to capture and why such a small adjustment would destroy its profitability.
Reserve data that you do not use to choose the rules. Test on that untouched period only after fixing the strategy. Once you inspect the result and adjust the system around it, that period becomes part of development, not an independent check. Keep a record of rejected versions too; the winning backtest should not erase the search that produced it.
Historical Trading Must Respect What Was Knowable
Audit every signal against the information available at its timestamp. A rule using a candle’s final closing price should not assume an entry before that close was known. For share strategies, check whether the historical universe includes companies that later disappeared, rather than only today’s survivors.
Also inspect simulated fills. If one candle touches both a stop and a profit target, its high and low alone do not reveal which came first. Use finer data where available, or flag the result as ambiguous rather than automatically awarding the profit.
These checks address a basic distinction: simulated trades are not executed trades. The CFTC’s hypothetical performance disclosure explicitly warns about hindsight and inaccurate treatment of liquidity effects. A smooth equity curve does not remove those weaknesses.
Trading Costs Can Consume the Entire Advantage
Rebuild the results using the costs of the instrument and account you would actually trade. Include spreads, commissions and plausible slippage, plus financing or borrowing charges where relevant. Avoid double counting a spread already captured in bid and ask execution prices.
Consider this hypothetical strategy, with identical position sizes throughout:
| Measure | Assumption or calculation |
|---|---|
| Win rate | 55% |
| Average gross winner | $100 |
| Average gross loser | $100 |
| Expected gross result per trade | (0.55 × $100) − (0.45 × $100) = $10 |
| Average total cost per completed trade | $12 |
| Expected net result per trade | −$2 |
The strategy wins more often than it loses, yet its assumed costs turn the expected result negative. The calculation matters more than the win rate. For the broader relationship between average returns and survival, see trading expectancy and risk of ruin.
A Signal Price Is Not an Executable Price
Separate signal generation from order execution in your records. Log when the signal appeared, when the order was submitted, what price was requested and what actually filled. Include rejected orders, partial fills and missed trades rather than analysing completed trades alone.
Order choice introduces trade-offs. For stocks, a market order does not guarantee an execution price; a limit order controls the acceptable price but does not guarantee a fill. A conventional stop order becomes a market order when triggered. These distinctions appear in the SEC’s explanation of order types. Check the corresponding terms for the product you trade.
Compare simulated and actual execution trade by trade. If your model earns only a small amount before costs, even modest differences deserve attention. Paper trading can help test the workflow, but do not treat its fills as proof that live orders will receive the same treatment.
The Strategy May Depend on Conditions That Have Changed
Write down the conditions your strategy needs. A hypothetical breakout system might require sustained movement after an entry; repeated reversals would undermine that premise. A strategy designed to trade reversals has a different dependency.
Split your evaluation by conditions defined without knowledge of the later outcome: volatility at entry, trading session, spread level or whether a scheduled announcement was approaching. Avoid inventing a filter simply because it removes yesterday’s losers. Any new filter needs fresh testing.
Set exclusion rules before the session begins. The guide to recognising poor trading conditions covers that decision separately. Here, the diagnostic question is whether live trades still match the conditions represented in your evidence.
Check Implementation Before Rewriting the Rules
Compare your actual trades with a parallel record of every valid strategy signal. If the rules required ten trades but you took six, exited two early and doubled the size of another, you have not run a clean live test of the original system.
Review operational errors as well: incorrect contract sizes, timezone mismatches, stale data and duplicate orders. Classify each discrepancy as a research assumption, execution issue or rule deviation. “Bad discipline” is too vague to repair.
Use a Controlled Review, Not an Emergency Redesign
Before deployment, define loss limits, review dates and conditions that require an immediate pause. Distinguish a technical fault from disappointing returns: broken order handling needs prompt action, while strategy evaluation needs enough relevant observations to support a decision.
Record those boundaries in a trading plan you can actually follow. When performance deteriorates, audit the data, costs, fills and adherence before changing parameters. If the evidence no longer supports the strategy, pause it. A backtest is a research result, not a debt the market owes you.