What “divergence” means in an MT4 Expert Advisor context
“Divergence” in MT4 Expert Advisors (EAs) generally means that two EA outcomes do not align. This can show up as different trade actions, different indicator-like values produced by the EA logic, different equity curves, or different signals derived from the same rules.
A key point is that divergence is an observation, not a single fixed cause. You first need to decide what is diverging:
- Two EAs of different code behaving differently on the same inputs.
- The same EA producing different results across runs.
- The EA’s live behavior differing from its backtest.
- The EA’s internal state changing differently than expected.
How divergence can be constructed from the EA’s mechanics
An EA is a program that reacts to inputs (price series, ticks, account state) and applies logic (rules, filters, risk routines). Divergence often appears when parts of that system are not identical.
Common sources of divergence include:
- Different inputs: If the EA uses parameters such as lookback length, timeframe, symbol mapping, or session filters, even small differences can change outcomes.
- Different execution timing: EAs may react to tick timing. If two environments feed price updates at different moments, the same rule can trigger at different times.
- State and control flow: Many EAs maintain internal state (for example, “already traded today” or “waiting for a condition”). If that state is initialized differently, divergence is expected.
- Assumptions in backtesting: A backtest uses historical data with its own modeling choices. Live trading uses real-time ticks and real fills. Even when the strategy rules are the same, the data and execution path can differ.
To reason about divergence, it helps to model the EA as: inputs → rule evaluation → orders/exposure updates → new state. If any link in that chain changes, divergence can follow.
Confirmation limits and why divergence can be misleading
Even when divergence is real, it is easy to interpret it incorrectly. Two psychological or analytical issues matter here:
Confirmation limits
If you only check cases where the EA “looks right,” you may miss divergence that occurs in other regimes. Confirmation limits also arise when you test too few situations (for example, only one market type or one volatility level). The result is that divergence appears to be an anomaly until you broaden verification.
Hindsight bias
After seeing that outcomes differed, it is tempting to believe the “cause” was obvious or inevitable in advance. Hindsight bias can make you construct explanations that fit the result rather than explanations that were defined before observing it.
This matters because divergence can come from multiple interacting factors (data, timing, costs, state). If you retrofit a single narrative explanation after the fact, you may overestimate how predictable the divergence was.
A concrete example (with explicit assumptions)
Assume an EA rule: “Open a position when a condition based on the last N bars is met.”
Assumptions:
- The EA uses N = 20 and evaluates on bar close.
- You run the EA twice, backtest A and backtest B, on the same symbol and timeframe.
Now introduce a divergence source:
- In backtest A, historical modeling feeds bar data so that the condition becomes true exactly at the bar close.
- In backtest B, due to data quality or different modeling behavior, the condition is evaluated slightly differently at the same bar close.
Under these assumptions, the EA may or may not enter on that bar. A single missed or added entry can later change exposure and state (for example, whether the EA is allowed to open more trades), leading to larger divergence downstream.
The important part is not the exact market outcome, but the logic: divergence can start from evaluation differences and then compound through state.
Material limitations, failure modes, and risks
Divergence is often a signal that your verification boundary is unclear. Some material limitations and failure modes to consider include:
- Overfitting and selection: An EA that was tuned to past behavior may diverge when conditions differ. - Execution and cost sensitivity: Results can change when spreads, commissions, or slippage differ from assumptions used during testing. - Data mismatch: Different historical feeds or tick data can produce different evaluation timing.