The usual version of this comparison is about judgement: a person reads context, an expert advisor reads a rule set. That framing is not wrong, but it is not measurable, and it hides the difference that actually decides outcomes. An EA follows its rules through a losing run. A person decides, in the middle of that run, whether the rules still apply.
We can measure the second half of that sentence on our own data.
But it depends on:
- Which rule you intervene by. A 3-loss rule stops 13 of 14 EAs. A 10% drawdown rule stops 1.
- When the losing run arrives. Four EAs are switched off while still in a loss, before their edge ever appears.
- What you are actually optimising. Sitting through a 775-day flat stretch is a cost, not a virtue, if you needed that capital.
- Whether the two are even separable. Most people who “trade manually” also run something automated, and the failure mode is the interaction between them.
Each of those is tested against our own data below.
What the data shows: the price of switching off
Every EA published here ships its full closed-trade list. Replaying those lists in order and applying a stop rule — switching off for good the moment the rule fires, because someone who quits after a bad run rarely restarts at the bottom — gives the cost of each intervention style.
| Stop rule | EAs switched off | Left sitting in a loss | Result given up | Median per EA |
|---|---|---|---|---|
| After 3 consecutive losses | 13 of 14 | 4 | 76,844.38 | 3,376.31 |
| After 5 consecutive losses | 7 of 14 | 3 | 63,717.06 | 1,713.32 |
| After 8 consecutive losses | 7 of 14 | 2 | 55,615.69 | 596.26 |
| Equity 5% below its peak | 10 of 14 | 4 | 68,346.36 | 2,855.88 |
| Equity 10% below its peak | 1 of 14 | 0 | 27,532.12 | 0 |
| Test conditions | |
|---|---|
| Experiment ID | EXP-INTERVENTION-COST-001 |
| MT5 build | 6090 |
| Environment | Replay of published closed-trade lists, 10,000 start balance per EA |
| Symbol / timeframe | One symbol and timeframe per EA |
| Period | 2019-01-01 – 2026-08-09, per-EA windows vary |
| Model | Every tick generated from M1 bars (12 EAs); every tick based on real ticks (2 EAs) |
| Last verified | 2026-08-14 |
The replay script and the per-EA rows are stored with experiment EXP-INTERVENTION-COST-001, so every line above can be recomputed from the same published trade lists. How we test covers the gate those records passed before they were listed.
Two things in that table matter more than the totals.
A tight stop rule fires early, not at the bottom. Under the 5-loss rule, Iridescence is switched off on 2019-01-14 — two weeks into a record that goes on to produce 9,403.46 — and the replay ends at -3.12. Lattice Weave stops on 2019-07-23 with 5.73 instead of 10,280.42. Tessera ends at -214.58 instead of 3,212.05. The losing streak that triggers the decision usually arrives before the edge has had room to show up at all.
A loose rule barely fires. The 10% drawdown rule stops exactly one EA — Cairn, whose published maximum drawdown is -14.52% — and leaves the other thirteen records untouched. That is the difference between a threshold picked from a record and one picked because the number felt safe.
Seven of the fourteen (Ballast, Beacon, Lanternfish, Orrery, Peregrine, Windrose and Zerqon) never hit five consecutive losses at all, and run to the end of their records under every streak rule tested. Whether an intervention rule costs you anything depends on which EA you attached it to — which is an argument for reading the record first, not for never intervening.
This is a replay of an owned record under a mechanical rule. It is not a live manual-trading record, and it does not measure discretionary skill — we do not have that data, and inventing it would be worse than not having it. What it measures is narrower and more useful: the cost of the specific reaction that separates the two approaches in practice.
EA vs manual: where the two actually differ
| Expert advisor | Discretionary trading | |
|---|---|---|
| Rule consistency | Identical on trade 1 and trade 400 | Varies with the last three outcomes |
| Reaction to a losing streak | Continues; the streak is in the record | Decides in the moment — the measurable gap above |
| Execution speed | Sub-second, on every signal | Limited by attention and screen time |
| Trade frequency it can sustain | Cairn closed 4,725 trades in this record | Not reproducible by hand at that rate |
| Context it can read | Only what is coded | News, regime change, anything visible |
| Failure mode | Keeps applying a rule set after it stops working | Abandons a rule set that still works |
| Where the record lives | Published, inspectable, recomputable | Usually reconstructed after the fact |
The two failure modes in that table are mirror images, and both are real. An EA has no idea its edge has decayed; it will keep placing trades with the same confidence it had in 2021. A discretionary operator notices decay quickly — and also “notices” it during ordinary losing streaks that are inside the documented behaviour. One error is expensive because it is slow; the other is expensive because it is fast.
Frequency is the other honest divide. Cairn’s record contains 4,725 closed trades over roughly five years. Nobody executes that by hand with consistent sizing at consistent moments, so for strategies that live at that frequency the comparison is not “which is better” but “which is possible”.
What to look for when deciding between them
1. The streak you can actually sit through. Take it from the record, not from imagination. The median worst losing streak across the catalogue is 6 and the longest is 14. If a run of 8 losses would make you intervene, and your chosen EA’s record contains one, you will intervene — and the replay above shows what that costs. Worst streak is the number to check before the profit factor.
2. A stop threshold that sits outside normal behaviour. The 10% drawdown rule works on this catalogue precisely because thirteen of fourteen records never reach it. A threshold inside the EA’s documented max drawdown is not a safety rule; it is a scheduled exit at the worst moment.
3. Whether you need the capital back on a schedule. Discipline is cheap to talk about and expensive to hold when the money is committed elsewhere. The median EA here spent 423 days at some point without making a new equity high — that is the real cost of the automated side, and it is covered in more detail in are expert advisors profitable.
4. Whether you want to change rules or only follow them. If your edge is reading context that no rule set captures, automating it removes the edge. If your edge is a rule you already trust and cannot execute consistently, automating it is the whole point.
Running both in practice
Most people end up doing both, and that is where the avoidable damage happens.
- Separate the accounts, or at least the accounting. An EA sizes positions against the balance it can see. A discretionary position that ties up margin silently changes the size of every automated trade after it, so the EA stops reproducing the record you chose it for.
- Give the EA its own magic number — the
MagicNumberinput in the EA’s properties dialog — so that manual closes never touch its positions, and read the Experts tab of the Terminal window rather than the chart when you want to know what it decided and why. Every EA published here ships its magic numbers as locked inputs for exactly this reason. - Know which switch you just used.
Tools > Options > Expert Advisors > Allow Algo Tradingis the terminal-wide setting; the AutoTrading toolbar button is the session one. Turning the second off pauses new entries and leaves open positions running, which is a different decision from closing the account down — and the replay above prices the second one. - Write the intervention rule down before you start — a specific drawdown percentage taken from the EA’s record, or a specific behavioural break such as a trade frequency that stops matching the documented rate. A rule invented during a losing run is a decision made at the worst possible moment.
- Decide what a manual override means. Turning AutoTrading off for a news release is a rule if you wrote it down beforehand and a reaction if you did not. Only the first one is repeatable, and only the first one can be evaluated afterwards.
If you want to change what the automated side does rather than only when it runs, the Builder is where the rule set is editable, because rewriting a condition is a different act from overriding it live.
Risk and drawdown reality
Three limits belong with the numbers:
A replay is not a person. The stop rules modelled here are mechanical and terminal. A real operator might restart, halve the size, or move to another instrument. Treating “switch off” as permanent gives an upper bound on the cost of that reaction, not an average of what humans do.
Backtests bound the past. These are strategy tester records, not live accounts. Spread widening, slippage on news and swap all subtract from the same edge and none of them appear in a replay.
Sitting still is not automatically right. The article argues against unplanned interruption, not against stopping. An EA whose drawdown exceeds anything in its record, or whose trade frequency departs from the documented rate, is telling you something a rule set cannot notice about itself. There is no guarantee of any result in either approach, and past results do not carry forward.
Next steps
If you want to see the records behind these replays, browse the verified EAs — each page publishes the complete closed-trade list, drawdown and run manifest, so you can check the worst streak and the deepest drawdown before deciding what your own intervention rule should be.
If you would rather encode your discretionary rules than override an EA’s, start in the Builder, and read are expert advisors profitable first for what the same catalogue says about the waiting time you are signing up for.