17 Oct 2026 · Study: Backtest versus live evidence · Research note

Why an EA backtest is profitable but the live account loses

Huyen X. Vu · Quantex Research, AlgoTrade Pte. Ltd., Singapore · ORCID 0000-0002-2209-842X

1. Introduction

A backtest is a model of the past. A live account is the record. When the two disagree, the cause is usually not one large error but several small gaps between the model and the market. This note sets out the common causes, a checklist, and what we learned when we tested one of our own models.

The question

Why does an Expert Advisor (EA) that is profitable in the strategy tester lose, or earn much less, on a live account?

2. Results

Six common causes

1. Costs are modelled too kindly

Spread, slippage and commission are small on each trade, but they are paid on every trade. A model with many trades and a small average result per trade is the most exposed. A fixed or narrow spread in the tester, no commission and no slippage can turn a negative model positive on paper.

2. Tick data is too coarse

The MetaTrader 5 tester can build prices in several ways. In "1 minute OHLC" mode the tick sequence is built only from the OHLC prices of minute bars, which saves time but can mislead. "Real ticks" mode uses the broker's tick history when it is provided (MQL5 book). Models with tight stops or intrabar exits are the most sensitive.

3. Overfitting and a short sample

If rules and settings are tuned until the past looks good, the test measures the tuning, not the effect. Standard safeguards such as a simple hold-out "tend to be unreliable and inaccurate in the context of investment backtests" (Bailey, Borwein, López de Prado and Zhu). A short sample makes this worse: a few months may contain only one kind of market.

4. Regime change

Markets move between calm and volatile periods. A model that relies on large moves can look strong in a volatile year and flat or negative in a calm one. If the test covers one regime only, the live period may simply be a different market.

5. Execution latency

In the tester, orders fill at the modelled price. Live, there is a delay between signal, order and fill, which can change the fill price and, around news, cause requotes or rejections.

6. Data and server differences

Data feed, spread at rollover and news, commission, swap, minimum stop distance, contract size and server time all differ between data sources and account types. A backtest on one data source is evidence about that source only.

Case study 01: seven months of history was not enough

In Case study 01 (12 Oct 2026) we re-tested one of our own systematic gold (XAUUSD) models on a much longer history:

• Method. The model was rebuilt as a simulation; on the largest adverse events it matched about 69–77% of the real ones. The same rules ran on 21 months and on 52.8 months of one-minute data (about 1.56 million bars), with costs, and single rules were changed slightly to test sensitivity.
• The short sample overstated the result. An earlier study on about seven months looked much stronger than the longer tests.
• The result came mostly from one regime. On 21 months, 2025 was flat for every variant; almost all of the positive result came from 2026, a year of unusually large gold moves.
• Calm years behave differently. On 4.4 years the best variant kept a small positive effect on average, but 2023 was negative in every variant.
• The result was sensitive. One small rule change (minimum distance between entries set to $0.10) turned the best variant negative and roughly doubled its worst drawdown.
• What changed. An earlier risk setting was withdrawn (on this history it would have led to a drawdown of about 99%). Risk is now set from a weaker variant with a margin of safety, and calm years such as 2023 are the base case.

3. Discussion

Checklist: before you trust a backtest

1. Costs: spread at typical and worst-case levels; commission and swap included; slippage added.
2. Data: finest tick mode available; data gaps checked.
3. Sample length: several years and several regimes, not only the recent past.
4. Sensitivity: change each rule slightly. If the result breaks, the effect is weaker than it looks.
5. Out of sample: keep data the rules never saw, then forward test on new data.
6. Risk: calibrate on the weaker variants and the worst historical move, plus a margin of safety.
7. Execution: measure real fill prices and delays on a small demo or live account and compare with the tester.
8. Same conditions: test on the same data source, account type and server you will run on.
9. Record: keep a live record others can check.

4. Limitations

Limits of this note

The causes and checklist are general research guidance; the numbers come from one model in Case study 01. A simulation is a model, not a record.

Appendix

FAQ

Why is my EA profitable in the backtest but losing live?
Usually a mix: costs modelled too low, coarse tick data, settings tuned to the past, a test period from a different regime, slower live execution and different data conditions. Check each one with the list above.

How much history is enough for an EA backtest?
There is no fixed number. It should cover several years and both calm and volatile periods. In Case study 01, about seven months made a gold model look much stronger than tests on 21 and 52.8 months.

Does testing on real ticks make a backtest match live results?
It brings the price path closer to real history, but it does not model live latency or every cost. It reduces one gap; it does not remove the others.

Related notes

• Gold weekend gaps: how often, how big
• LLM trading agents vs buy-and-hold: what the 2025-26 benchmarks show

Declarations

• Author: Huyen X. Vu, Quantex Research, AlgoTrade Pte. Ltd., Singapore (ORCID 0000-0002-2209-842X).
• Data availability: Public sources are listed under Sources. Results for our own models come from internal records that are not public; summary figures are shown in the note.
• Use of AI tools: Data processing, analysis code and drafting were assisted by AI tools (Claude, Anthropic). The author checked the results and is responsible for the content.
• Conflict of interest: The author is the founder of AlgoTrade Pte. Ltd., which develops and runs systematic trading models.
• Funding: No external funding.
• Peer review: Not yet peer reviewed. Comments are welcome at info@quantex.com.sg.
• How to cite: Vu, H. X. (2026). Why an EA backtest is profitable but the live account loses. Quantex Research, AlgoTrade Pte. Ltd. https://quantex.com.sg/research/2026-10-17-backtest-vs-live/

All notes and case studies

For education and information only. Not investment advice. See our Disclaimer.