Two theses, and the method behind them

The whole thing, in five sentences

A prediction market is a market where you buy a contract that pays $1 if some event happens and nothing if it does not, so a contract trading at 63¢ means the crowd thinks there is roughly a 63% chance — the price is the forecast. I collected these markets moving, second by second, and asked a single question: does a market have a hidden mood that turns busy shortly before its price starts moving, or is it simply that price moves cause more price moves? If the first were true you could see activity coming, which would be useful and slightly surprising. I found the second: the model that assumes a hidden mood detects busy spells but never anticipates them, and a model where each price move simply makes the next one more likely predicts better. Along the way the markets did something I did not expect — as a market approaches its deadline it goes quieter, not busier.

Everything below is that story in more detail. Terms are defined where they first appear, and collected in the glossary.

The MSc thesis process, results and limits →

A result that goes against you is a credential

Three of the four findings above went against what I hoped to show, which is not how most research reads. The reason is an incentive: when a result comes out flat, the easy path is to say nothing, adjust the measurement, and try again until something looks convincing. Nobody sees the versions that failed, so nobody can tell the difference between a real finding and the one attempt out of twenty that happened to look good.

Writing down what counts as success before looking at the answer, and committing it with a timestamp, closes that path. If I move the goalposts afterwards, the history shows it. That is the entire value of the practice, and it is why I would rather show you a null result I pre-registered than a strong result I did not.

How I keep myself honest

The reason to trust any number above is the protocol that produced it, so here is the protocol — including where it falls short.

• Decide what counts as success before looking. The MSc work carries 14 pre-registered components, each committed to version control before the corresponding result, so the order is provable. One amendment exists and is disclosed: the original scoring rule failed its own simulated false-alarm check, so an already-registered alternative replaced it — before any real number was computed.
• Make the test prove itself first. Several tests had to demonstrate, on simulated data with a known answer, that they could find a real effect and return nothing when there was nothing there — before being pointed at real data.
• Allow "no effect" as an outcome in advance. Writing "no answer" into the specification as an acceptable result is what makes reporting one afterwards meaningful rather than merely disappointing.
• Count the right thing. Uncertainty is estimated by resampling whole markets, not individual trades — see clustered bootstrap — because trades within one market are not independent, and treating them as though they were would make every result look more certain than it is.
• Never test on what you trained on. Leave-one-out as the standard, time-ordered splits where time matters, and any deviation disclosed where it is used.
• Correct for asking many questions. Test twenty things and one will look significant by luck; the correction accounts for that before any result is called.