Nearly 80% for Two Days—What Happened Over Five Years?

The most dangerous number in quantitative research is not always a bad result.

Sometimes it is a result that is:

so good that you desperately want to believe it.

We recently encountered exactly that while studying second-level microstructure in Polymarket BTC 5-minute markets.

In a very short recent sample, some combinations of an early directional signal and a microstructure confirmation condition produced win rates approaching 70%–80%.

If this were a marketing screenshot, the research could have ended there:

80% WIN RATE
NEW EDGE FOUND

We did the opposite.

We moved the same type of relationship into nearly five years of continuous historical data.

The result was much more typical—and much more useful.

1. The “Magic Pattern” in a Small Sample

Why are short samples so dangerous?

Because a local period can accidentally be perfect for one feature.

For example:

volatility may be unusually consistent;
directional trends may persist;
one microstructure behavior may dominate;
participant behavior may temporarily align.

A simple condition can then look extraordinary.

The problem is:

You do not know whether you discovered a durable structure or merely discovered the last few days.

And once a beautiful result is visible, the human brain becomes very good at inventing reasons for why it “must” make sense.

That is where research becomes most dangerous.

2. We Did Not Try to Optimize the 80%

The tempting next step would have been:

Tune the recent sample a little more and see whether 80% can become 82% or 85%.

We went in the opposite direction.

We stopped tuning the recent sample and expanded the research across years of history.

That required rebuilding a much larger second-level data environment and testing the same general relationship across multiple market regimes.

The spectacular short-sample win rate did not survive.

Some candidate relationships remained above random and retained positive standardized historical results.

But overall they moved back into:

the mid-50% range.

That is a completely different story from 70%–80%.

3. That Is Not a Failure

It is easy to say:

“If 80% became roughly 56%, the idea failed.”

We think the opposite.

This is the point where the research became meaningful.

A short sample can tell you:

“A strong relationship appeared recently.”

A long history can begin to answer:

“Does that relationship exist across time?”

Those are different questions.

After multi-year validation, the microstructure effect did not disappear completely.

It still contained information.

But it was nowhere near strong or stable enough to justify being treated as a standalone production strategy.

That is a much more realistic conclusion.

4. Why Multi-Year Validation Matters More Than One Headline Win Rate

Even when the full-history result is positive, total win rate is not enough.

We care more about questions such as:

Does the relationship survive different years?
Is there an obvious failure period?
Is most of the result concentrated in a small number of regimes?
Does drawdown become unacceptable in certain environments?
Does recent behavior differ materially from earlier history?

In this research, one pattern was especially important:

There was information across the full sample, but not every year behaved equally well.

That suggests the feature is more likely:

conditional information

than:

a universal standalone rule.

5. “Contains Information” Is Not the Same as “Ready for Production”

A feature can be:

statistically related to the target;
slightly above random across full history;
very strong in some periods;

and still be:

not ready for a production trading system.

Production rules require more than:

correlation exists

They require something closer to:

time stability
robustness
low parameter sensitivity
reasonable behavior across regimes
continued value when combined with the existing system

So our current treatment is simple:

keep the research value, do not freeze it into a production rule.

6. Why We Think Failed Research Is Worth Publishing

Most quantitative content shows only what survived.

You rarely see:

the idea that looked amazing initially;
the rule that failed over longer history;
the parameter that only worked recently;
the feature that contained information but was not stable enough.

Showing only the winners creates an illusion:

Good strategies are discovered cleanly and directly.

Real research is usually closer to:

observe a pattern
↓
form a hypothesis
↓
get a strong local result
↓
expand the sample
↓
watch the edge shrink
↓
redefine what the feature is actually useful for

This study is a good example of that process.

7. What We Actually Kept

After the five-year validation, we did not keep an “80% strategy.”

We kept three more useful conclusions.

First

Short-horizon market microstructure does contain measurable information.

Second

That information is extremely easy to overestimate in a small sample.

Third

It may be more useful as a confirmation layer around an existing decision than as a standalone directional strategy.

Those lessons are more valuable than a short-term screenshot.

Conclusion

If a Polymarket bot research result approaches 80% over two days, the best next step may not be deployment.

It may be:

Try to break it first.

Expand the history.

Split by time.

Check for leakage.

Find the weak regimes.

If the relationship survives all of that, then keep studying it.

This is less exciting than a headline win rate.

But it helps reduce one of the most dangerous mistakes in quantitative research:

confusing a recent accident with a durable edge.

Disclaimer: This article discusses a research-validation process. Local win rates and multi-year results do not represent actual account returns, and no reproducible strategy condition, threshold, parameter, or signal combination is disclosed.