Backtests & Track Records: What Historical Results Really Show

A backtest answers a conditional question.

A backtest applies a defined set of trading rules to historical data under specified assumptions. It asks: How would this version of the strategy have performed in this dataset and execution model? It shows no actual account performance achieved and alone does not provide a reliable promise for future profits.

This makes it valuable: rules can be controlled, mistakes discovered, costs examined and unsuitable ideas discarded before real capital is used. The transition from a retrospective calculation to a forward-looking assertion becomes problematic if its presuppositions are concealed.

“Data-based” initially only means that data has been used. Whether the data were available in a timely manner, the rule was only created after consideration of the result and the execution was realistically modeled, remains open.

Hindsight and live decisions use different information.

A retrospective analysis does not necessarily use future information. A sound test reconstructs what was known at each point and generates decisions in chronological order. For example, a signal based on the daily closing price cannot also be filled at an earlier intraday price if the information required to generate it was only available later.

The difference still remains: the researcher can know the course, even if the program code only reads past prices. Knowledge of a later crisis can already influence the choice of instrument, filter or test period. Therefore, “no future access in the code” is not sufficient as proof against all retrospective distortions.

Predictive validity requires additional assumptions: The context under investigation must continue, the available data must be comparable and the strategy must remain actually executable. An interest rate regime, market structure, competitor or broker conditions may change. The future is not another line of the same immutable experimental setup.

What a reproducible test requires.

1. Freeze the strategy: Record the version, parameters, signal time, position size, exit and risk limits before the evaluation.
2. Document the data: Record the source, period, time zone, Bid/Ask/Last prices, resolution, gaps and corrections. For equities, also account for the companies available at the time and corporate actions.
3. Describe execution: Separate signal price and tradable price; name order type, latency, spread, slippage and partial filling assumptions.
4. Define the account model: Capital, leverage, margin, currency conversion, financing, commission and possible forced closures.
5. Record the selection process: Not only save the winner, but also examined variants and the pre-selected rating size.
6. Result completely show: Export trades, open positions, equity, balance, drawdown, costs and error journal.
7. Independent verification: Repeat the same rules with traceable data; explain and document deviations.

A screenshot does not reconstruct data quality or selection process. A second evaluation of the same file checks calculations, but does not automatically confirm the authenticity or completeness of the initial data. A file hash protects the assignment of a version; it does not prove that the trading data is true.

MetaTrader: the test mode defines the simulation.

Among other things, MT5 distinguishes real ticks, generated ticks, 1-minute OHLC and open-prices tests. Suitability depends on whether a strategy responds to movements within a candle. A test only on opening courses cannot test an intrabar-dependent logic equally. The tester offers deceleration models; the mode without deceleration expressly maps idealized conditions. In the case of profit calculation in pips, swap, commission and margin audit are eliminated according to the documentation. [1]

Real ticks are closer to the observed course, not a guarantee of completeness. The MT5 documentation also describes generated ticks as a replacement when tick data is missing from an existing minute candle. For real ticks, the spread can change within the minute; for generated ticks, the corresponding minute spread is used. In addition, a last chart does not replace the bid/ask prices that are relevant for trading. [2]

High modeling quality in itself says nothing about economic viability, independent choice or future profitability. Even good price data does not automatically know your real position in an order queue, every rejected order ticket or the market impact of a large position.

MT4 and MT5 should not be equated via seemingly identical report labels. Test data, models and trade mechanics. The MT4-/MT5-Strategy Tester GuideBacktests and Track Records Describes the operation; this page assesses the significance. Additional: Execution quality, Contract data and Cost models.

How a convincing curve can rest on weak evidence.

Bailey, Borwein, López de Prado and Zhu are investigating selection bias through backtest optimisation. Their approach considers, among other things, the relationship between in-sample selection and out-of-sample performance. A simple withheld dataset does not automatically answer how much previous multiple attempts affected the selection. Their research is not a blanket diagnosis of each EA; it justifies why the entire selection process must be part of the proof. [3]

Robust results don’t have to look spectacular. A wide range of similar results for neighbouring parameters should be assessed differently than a single narrow peak value. Even such an area is not proof of the future: the entire parameter family can depend on the same historical market phase.

Selection lab: the winner among many random trials.

All candidates in this teaching model have the same expected gross value of zero. Each has 60 randomly produced historical results and 60 other independent results. The selection is based solely on the highest historical total profit. The second half does not affect the selection.

Own synthetic simulation: per step independently +10 or −10 money with equal probability; constant size, no costs, no real rates. The starting value makes the experiment reproducible. For the same starting value, smaller candidate sets are subsets of larger sets. No testing of real strategies, no PBO calculation, and no probability of future success.

The selected candidate can win or lose in the second section. Both are possible. A single further positive result therefore proves no advantage. Consciously change the starting value and compare several complete experiments. Keeping only pleasant experiments would re-create the same selection error.

Evidence lab: what an attractive balance curve can hide.

An invented account starts with 10,000 money. It closes a gross profit of 100 in twelve steps each, while another position temporarily carries an open loss of 2,000. The lines show the complete artificial course, including the starting value.

No deposits/payments, no interest and no margin calculation. The costs are a flat deduction per closed trade and are not intended to replicate a brokerage fee. Open P/L is additionally taken into account without deducting the costs again. Drawdown is calculated from the displayed model times including starting value; unobserved intermediate movements are missing.

That the same strategy could produce other curves is not shown by this fixed example. This would require further paths and verifiable assumptions. The laboratory demonstrates the effect of lack of information on the observation of a given course.

Interpreting performance metrics.

The MT5 test report defines profit factor, various balance/equity drawdowns and Sharpe ratio. The concrete implementations and reference values must be considered in the comparison. [4] Measurements should be considered net, over comparable periods and with reported open positions.

Deposits are not a trading profit. With changing capital inflows, simple change in account value, time-weighted return and money-weighted return differ. Even the annualization of short test periods can create a false impression of stability. An accurately calculated number cannot improve a weak data base.

From historical test to prospective observation.

1. In-sample: Data for development and parameter selection.
2. Out-of-sample: Data that did not affect this selection. After sighting and adjustment, they are no longer an untouched test for the new version.
3. Walk-Forward: Repeatedly calibrate on a previous section and check the following section according to the fixed procedure. Define window, selection criterion and adjustment frequency in advance.
4. Prospective Demo-Forward: Set version and parameters before the start; observe new data only in the course. Technical operation becomes visible, real execution remains limited.
5. Live observation: Fully document actual transactions and costs. A live account does not eliminate account selection or missing open losses from the presentation.

Definition limit in MT5: The “forward” section of the strategy tester is a later part of the selected historical period. It is not automatically a future live or demo period. Its independence depends on whether it has already been incorporated into the development. [1]

None of these methods alone provide a seal of quality. For rare extreme events, even a long profitable phase may contain few observations. Many closely related trades are not as much independent evidence. A rigid minimum number or universal drawdown boundary would be misleading without a strategy and data context.

From misunderstanding to deliberate misrepresentation.

A methodological error can arise from ignorance. A shortened presentation may be negligent. If unsuccessful variants, open losses or unrealistic assumptions are deliberately concealed, the presentation can create a wrong impression. The intention cannot be determined from the curve alone. Verifiable data and omissions are therefore critically examined.

As an example of this distinction, the US-SEC treats backtest results in the context of its rules for investment advisor marketing as hypothetical performance. Their materials emphasize the importance of assumptions, risks and limitations. This is a US-specific regulatory context and not a blanket legal statement for every EA seller in Europe. [5]

The next useful step: request the evidence.

Download the checklist (markdown)

Ask providers for the complete test protocol, strategy version and parameters, data source, all cost assumptions, the selection process, all open positions and a traceable independent test phase. Ask for changes between backtest and live account, and full transaction data instead of screenshots only.

Decision rule: A lack of proof is not evidence of fraud. However, it limits what conclusions you can responsibly draw. A spectacular result without a reproducible basis deserves less trust than a more modest, cleanly documented investigation.

The references combine technical verification tasks; no implemented data integration is claimed. Additional: Exporting reports, Understanding Expert Advisors and Secure evidence and protect consumers.

Sources and limits of the evidence.

Research status 01.10.2026. Manufacturer sources explain tester functions and report definitions; the scientific work examines selection bias. The critical classification and calculation examples are our analytical workup. The teaching models do not use real market data and do not evaluate a specific product.

1. MetaTrader 5: Strategy Testing — Modes, delays, pip calculation and historical forward range.
2. MetaTrader 5: Real and Generated Ticks — Spread, bid/ask and replacement of ticks produced.
3. Bailey, Borwein, López de Prado and Zhu: The Probability of Backtest Overfitting — scientific methodology; the selection laboratory does not implement their PBO procedures.
4. MetaTrader 5: Testing Report — Key figures and calculation limits.
5. SEC: Investment Adviser Marketing — US-specific treatment of hypothetical performance.