Beating random isn't enough: we pitted the engine against three dumb strategies

A hit rate means nothing on its own. "68% correct signals," fine — but compared to what? If the market rises 55% of the time, predicting "it goes up" every time already gives 55%. The only question that matters: does the engine do better than dumb strategies on exactly the same situations?

Three opponents, a fair match

We defined three naive strategies, then judged each on the same assets, the same dates, the same horizons as the engine, with the same resolution rule. No home advantage.

• Random (50%): a coin flip.
• Drift ("always up"): predict "it rises," always. It captures the natural upward bias of markets.
• Naive 20-day momentum: predict it continues in the direction of the last 20 days.

As with the previous analysis, the definitions were written before seeing the results and published whatever the figure.

The verdict, unfiltered

Opponent	Engine's edge	Significant?
Random 50%	+16.7 pts	yes (p = 0.003)
Naive 20d momentum	+16.0 pts	yes (p = 0.006)
Drift (always up)	+2.7 pts	no (p = 0.39)

Two good pieces of news and one honest one.

We beat random and naive momentum, clearly and stably across two independent backtest runs. "Take the sign of the last 20 days" does not reproduce the engine — which matters, because many "systems" are nothing more than disguised momentum.

We don't (yet) beat drift. Over the measured period, 64% of resolved decisions ended up, and "always up" does almost as well as we do. Two factual nuances, not excuses: the test is under-powered (only 12 cases where the engine and drift disagree), and above all drift says nothing about risk — it's always exposed, on everything, with no sizing. It's the volatility-scaled P&L and the upcoming paper trading that will settle this point in real money. But on the pure-direction criterion, the honest claim today is: "beating drift is not demonstrated."

Why show a weakness

Because pre-registration only means something if we also publish what bothers us. Anyone can now read, on the Truth Table, the engine's edge against these three baselines, with the statistical test (exact McNemar) and the exact count of discordant cases. The full report is here.

An engine that beats random and momentum but admits it hasn't yet beaten drift is less catchy than a trumpeted "68%." It's also the only version you can trust.