The Renegade Rule. A stranger explained the physics and it changed the bot the same day.

A reader left a comment on our graveyard post that changed how the bot trades within about four hours. This is what he said, what we measured, and what it cost us to have been wrong for a month.

The comment

Two claims. First, that the CF Benchmarks index our crypto markets settle on is a martingale by construction, meaning the best forecast of its next value is always its current value. Second, that the relationship between a binary contract's price and its order book is more nuanced than spot or futures, because the contract's sensitivity to the underlying explodes near the strike as time decays toward settlement.

The first claim explains our graveyard. Every directional strategy we tested lost money, and we had filed that under "hard problem." It is not a hard problem. It is arithmetic. If the underlying is a martingale then no forecast of its direction has positive expected value, and every one of those experiments was doomed before it was written. That is a more useful thing to know than a pile of individual failures.

The second claim is the tradeable one.

The idea

A fifteen minute market asks whether an index will be above a strike at settlement. The honest way to price it is not to ask which way the market is moving. It is to ask a physical question. How far is the index from the strike right now, and how far can it possibly travel in the time remaining?

If the index sits far from the strike with sixty seconds left, the outcome is nearly determined regardless of what happens next. If it sits on the strike with the same sixty seconds left, it is a coin flip. Both markets can be quoted at eighty cents by a crowd watching price rather than distance, and they are not remotely the same bet.

We had been measuring the wrong variable. Our existing filter watched how often a market's price crossed its own recent average, which is a shadow of volatility rather than volatility itself.

The measurement

We sorted 2,366 of our own historical entries by distance to strike, expressed in units of how far the index could still travel before settlement. Call that number z. A z of one means the strike is exactly one standard deviation of remaining movement away. All entries are the same strategy, the same price band, the same windows.

Sixty four percent of our historical trades sat in the first bucket. They produced about three percent of the profit. Put the other way, roughly a third of our trades earned nearly all of the money, and we had no way of telling them apart until a stranger explained the physics.

By pond, the effect is strongest where the strikes sit closest to the money:

What we did about it

Both live lanes now refuse entries where the strike is still within one standard deviation of remaining movement. The filter is named the Renegade Rule, after the reader who explained it.

It went live before clearing our usual two standard error bar, which we do not normally allow. The reason is that this filter can only ever remove trades. It cannot cause a loss that would not otherwise have happened, so its downside is opportunity cost rather than risk. A paper twin runs the identical strategy without the filter beside it, so the honest verdict is still being measured rather than assumed.

We also started recording the settlement index itself. It is free, official, and was one subscription line away on a websocket this lab already had open. For a month we had been inferring settlement from spot prices, which is not the number these markets settle on.

What this does not prove

The study is in sample. These are the same trades that produced the rule, and a filter fitted to its own history will always look good. The forward test is the only one that counts and it started today.

The alt coin history is short. Only BTC has three weeks of reconstructable spot data, so the per pond numbers outside BTC rest on two days.

The measure conflates two things. Distance to strike divides by the square root of time remaining, so an entry made early in a window scores lower than the same distance made late. Some of what this rule captures may be a preference for later entries rather than for safer ones. Untangling those is the next study.

And two of our own lanes disagree with it. The favorite maker paper lane and the queue twin both made most of their money in the bucket the rule discards. Any rule that improves three lanes and hurts two is not finished.

The part worth keeping

The most valuable thing in the lab this week was not a strategy. It was a stranger explaining that we had been measuring a shadow of the real variable. We published our failures, someone read them, and the correction came back for free.

That is the whole argument for doing this in public.