Does an Algorithmic Rule Beat Random Entry? 4 of 5 Rules Lost to a Coin Flip

Every algorithmic trading article starts from the same unstated assumption: that a mechanical rule is better than no rule, because it is disciplined and repeatable. Nobody checks it. So I checked it. I built the dumbest possible opponent — a coin flip — gave it 2,000 chances, and made five well-known entry rules play against it on the same 16,926 EURUSD 5-minute bars from June 3 to August 26, 2026.

How do you make a coin flip a fair opponent?

One random trade proves nothing. The trick is to run enough of them to see the whole distribution of luck, then ask where the rule falls inside it.

• 2,000 independent runs. Each run takes 300 trades.
• Each trade: a random bar out of the 16,926, a random direction, held exactly 24 bars — two hours — then measured in pips. No stop, no target, no spread, for the coin flip and for the rules alike.
• Deterministic generator. Seeded once so the whole study reruns identically. Nothing here is cherry-picked from a lucky pass.
• Rules scaled to 300 trades by their average pips per trade, so a rule that fires 460 times is not flattered against one that fires 281 times.

What does pure luck actually look like?

Two numbers deserve attention before we go further. First, 45.9% of random portfolios finished in profit. Almost half. Second, the best of them made 628 pips over 300 trades on nothing but dice. That is the real reason so many strategies look brilliant in a backtest and die in an account: the range of luck is enormous, and one backtest is one sample from it. This is the same problem I described in my honest answer on whether AI can predict forex — the hard part is never producing a good-looking result, it is proving the result is not noise.

Which rules beat the dice?

Read the last column as: what share of coin-flip portfolios this rule beat. Donchian breakouts and naked Tenkan/Kijun crosses beat 8.7% of them. Put plainly, if you had thrown dice instead, you would have done better nine times out of ten.

And look at the win rates. Every losing rule has a win rate below the coin flip’s 47.5%. That is not bad luck, it is structural. These rules all fire after a move has already happened, which means they systematically buy the top of a small push and sell the bottom of a small drop. The dice have no such bias, which is exactly why they are hard to beat. My moving average crossover test found the same shape from a different angle, and my Stochastic overbought study found it a third time.

Why did the one winner win?

The rule that cleared the 95th percentile is not a better trigger. It is the same Tenkan/Kijun cross that finished at the 8.7th percentile — with one condition added: only take it when price is already on that side of the Ichimoku cloud. That single filter moved it from -276 to +330 pips per 300 trades, a swing of over 600.

This is the pattern I keep running into. A fast trigger has no opinion about context and fires constantly in chop. A slow filter has an opinion but is late as an entry. Slow licensing fast works; I broke the mechanism down properly in my Ichimoku versus plain trend following comparison, and the reverse combination failed in my ADX against Supertrend test. Four separate studies, one direction.

But I will not oversell it either. 96.4th percentile means 3.6% of pure coin flips still did better. That is a real edge, not a certain one, on one pair over one quarter.

What should you take from this?

• A rule is not evidence. Being mechanical only removes emotion. It does not create an edge, and four of five popular rules here had a negative one.
• Ask for the random baseline. When a strategy is sold on “+400 pips backtested,” the only useful question is what 2,000 coin flips scored on the same data. If nobody can tell you, the number means nothing.
• Frequency is a warning sign. The two worst rules produced the most signals — 460 and 427. The winner produced the fewest, 281.
• Filters beat triggers. Every positive result in this series came from removing trades, never from finding a smarter entry.
• Win rate is not the score. The coin flip had a better win rate than four rules that all lost money to it anyway.

Frequently asked questions

Does this mean technical indicators do not work?

It means these five entry triggers, measured naked over a fixed two hour hold on EURUSD M5, mostly did not. The one that worked was an indicator too. The distinction that survived the test is not indicator versus no indicator, it is trigger versus filter — indicators used to say “not now” earned their place, indicators used to say “enter now” did not. My ADX above 25 test is the cleanest example of the filter case.

Why 300 trades per random run?

Because it is the same order of magnitude as the rules produced — between 281 and 460 signals over the quarter. Comparing a 300-trade rule against a 10,000-trade random baseline would understate how wide luck really is at realistic sample sizes.

Would adding a stop loss change the ranking?

It changes the totals, not usually the order, because a stop is applied to winners and losers alike. I measured that separately in my fixed stop versus ATR stop test, where the stop choice moved results a long way but never rescued a negative entry.

Is 46.6% a good win rate?

On this measurement, yes. There is no target and no stop, so wins and losses are whatever the market gave in two hours. A 46.6% win rate with +1.10 average pips beats a 55% win rate with a negative average every time.

Can I trust a backtest at all after this?

You can trust a backtest that reports its baseline, its trade count and the range it could have produced by chance. You cannot trust an equity curve with no context. That is the entire point of running the dice first.