Algo Backtesting: A Reproducible Forex and Futures Workflow

A backtest runs specified rules on historical inputs under a defined simulation model. It can reveal how that implementation behaved in those conditions. It does not reproduce every possible fill, establish a future loss ceiling or prove permission for a prop-firm account. The useful output is a reproducible record, including unsuccessful tests.

1. Freeze the question and the implementation

Write the instrument, intended decision time, entry/exit conditions, sizing, session and failure behaviour before selecting results. Save the exact source or executable version, input file and dependencies. A moving-average or RSI example is a specification to test, not evidence of an edge.

Define which data may be known at each decision. A bar's closing price is unavailable at its opening. Include the publication time of news and revised data where relevant. Record the baseline and the decision criteria before examining the final evaluation period.

2. Audit the data and contract assumptions

Record feed, coverage, timestamps, timezone/DST, bid/ask availability and missing intervals. Investigate anomalies before correcting or excluding them, retain the original and log the change. Price jumps can be genuine. Tick data, generated ticks and OHLC bars are different inputs; a 99% or 99.9% quality label is not a probability that a result will hold.

For futures, record the actual contract, expiry, roll dates, session calendar, tick size/value and any continuous-series adjustment. A back-adjusted series is not itself a list of prices at which the expiring contracts could be traded. Model a defined roll method and its costs. For forex, document symbol specifications, financing and the broker's actual session/holiday handling.

3. Configure and inspect the simulator

Choose the tester matching the actual program and record modelling mode, account currency, deposit, leverage, volume steps, costs and execution assumptions. In MT5, recorded real ticks and generated every-tick modes are distinct; the tester also provides execution-delay settings. Neither mode duplicates all queue, liquidity or broker behaviour.

Run a small sanity test before optimisation. Inspect entry timestamps, fills, stops, quantities and account calculations against the specification. Include spread, commission, swaps and execution costs once. If an assumed gross result is 20 pips and separate spread/commission costs are 1.5 and 0.5 pips, the remainder is 18 pips before other costs, not automatically a loss.

4. Separate selection from evaluation

Select or optimise parameters only on the designated development data. Reserve a genuinely unused period for evaluation. In a walk-forward design, repeat chronological selection and evaluation windows and save every chosen parameter set. The MT5 Forward option is a held-out historical period; it is distinct from a later demo run.

Report all variants searched and any changes made after seeing evaluation results. Reusing a holdout to select the next winner makes it part of development. Sensitivity and resampling tests can expose dependence on assumptions, but their results are conditional on the design; randomly shuffling trades can remove real dependence and does not prove the worst future drawdown.

5. Read the report with the right definitions

Record net and gross results, costs, activity, average outcomes, balance/equity drawdowns and their currency or percentage definitions. Profit factor divides gross profit by the magnitude of gross loss, with a non-zero denominator. Recovery factor relates profit to maximum drawdown, not elapsed recovery time. Sharpe comparisons need consistent return sampling, reference rate and annualisation.

There is no universal PF, Sharpe, win-rate, number of years or number of trades that certifies a strategy. Sample dependence, selection and the uncertainty of the measured result matter. An observed peak-to-trough drawdown is also distinct from a provider's initial-balance or trailing loss rule; replay the exact programme/phase calculation, open exposure, costs and reset time.

6. Save a test package and state what remains unknown

Keep data provenance, the exact implementation/settings, all search choices, run configuration, reports and rejected results together. A rerun should reproduce the stated result within documented platform/data limitations. If it does not, investigate the mismatch instead of replacing the report with a more favourable run.

Operational checks and later authorised demo observations are separate evidence. A passing test is not an instruction to risk capital or scale. The JPTC EA Hub offers downloadable files: check each actual file's documentation and report. The separately hosted Algo requires no customer installation/VPS; its company-published monthly broker figures do not audit another strategy or establish its future results.

Sources checked 16 October 2026: Strategy Tester · MT5 test report · EA inputs and permissions · Sharpe (1994) · Data leakage.

Futures roll methodology (Campbell/CME)

Does Your Trading Algorithm Work? Read the Evidence · Test a Trading Algorithm: The Deployment Readiness Checklist · JPTC EA Hub · JPTC Algo

Broker-statement images of the separately hosted JPTC Algo

These service images do not establish another algorithm’s results, execution or provider permission.

@jptradingcapital on Instagram →

FAQ