Field Study 02  ·  2026.07.26

Chasing supernovas: quant finance research

A stock opens way up. Does it keep climbing, or give it back? I measured 3,083 of these across 45 companies. The answer turned out to depend on one thing only — and then I spent a day trying to prove that wrong, and failed.
3,083 jumps45 companies2 industries3.5 years
Instruments
Python/pandas/statsmodels/hmmlearn/vectorbt
Methods
Hidden Markov models/confidence intervals/statistical power/data quarantine
01 / The Question

Some mornings a stock opens way above where it closed.

Usually because something happened overnight — a drug trial worked, a contract landed, a rumor spread. Traders call the gap between yesterday's close and today's open exactly that: a gap. The question every trader faces at 9:30am is simple to ask and hard to answer: is it too late to get in?

So I asked it as a measurement instead of an opinion. Take every day a stock opened at least 3% higher than it closed. Check whether it ended the day above where it opened. Call that "kept going." Then look for anything that predicts it.

I ran this on two industries that have nothing to do with each other: small biotech companies, and small AI and quantum-computing companies. Different businesses, different news, different investors. If something showed up in both, it probably wasn't a coincidence.

02 / The Finding

How many people showed up. That's it.

Every stock has a normal amount of daily trading. Some days almost nobody trades it; some days there's a stampede. I measured each jump day against that stock's own recent average — call it how busy the day was.

Sort the jumps by busyness and the result is almost embarrassingly clean:

Biotech — 965 jump days across 24 companies
0%20%40%60%80%TYPICAL DAY 39.6%31%quiet day<1.5x · 599 days46%a bit busy1.5-2x · 144 days50%busy2-3x · 101 days69%very busy3-5x · 64 days67%frantic5x+ · 57 daysHOW BUSY THE DAY WAS →

On the quietest days, the stock kept climbing about 31% of the time. On the most frantic, about 67%. That's a swing of roughly 36 points, driven by nothing but how many people were in the room.

The thin white line on each bar is honesty about sample size — it shows the range the true number could plausibly sit in. Notice the last two bars: their ranges overlap heavily. Which means the climb is real, but it flattens out near the top rather than continuing forever. Beyond "very busy," more busy doesn't buy you more.

AI & quantum — 2118 jump days across 21 companies
0%20%40%60%80%TYPICAL DAY 39.3%29%quiet day<1.5x · 1350 days49%a bit busy1.5-2x · 243 days56%busy2-3x · 215 days60%very busy3-5x · 148 days68%frantic5x+ · 162 daysHOW BUSY THE DAY WAS →

Two industries with no overlap in companies, news, or business model produced nearly the same curve.

That's the part worth caring about. If only biotech had shown this, it might be a quirk of how drug news travels. Showing up twice, independently, means it's probably something about crowds and small companies in general.

03 / Trying To Break It

Does the mood of the market change the answer?

Here's the obvious objection. Markets go through stretches — sleepy months, wild months. Maybe this whole pattern is really just "2023 was crazy," dressed up as a finding.

To test that I needed to label every day as calm or jumpy. But I didn't want to draw those lines myself, because I'd unconsciously draw them where they'd flatter the result. So I used a hidden Markov model — which sounds intimidating and isn't.

Imagine watching someone's mood through a wall. You can't see them, but you can hear how loudly they're moving around. Over weeks you'd notice the noise comes in stretches: quiet stretches and loud stretches, each lasting a while before flipping. A hidden Markov model does that arithmetic. It finds the stretches on its own, from the data, with nobody telling it where to put the boundaries.

One rule mattered more than any other: no peeking. It's very easy to build a model that labels last March using what happened in April, then congratulate yourself on how well it works. Every label here was computed using only what was knowable on that day.

Then I split all 2,977 labeled jump days by mood and drew both curves on the same axis:

Same measurement, split by market mood
20%40%60%80%CALM STRETCHESJUMPY STRETCHESquiet daya bit busybusyvery busyfrantic

They sit on top of each other. Calm stretches: 39.4%. Jumpy stretches: 39.8%. A gap of four-tenths of one percentage point.

Before running it I worked out how big a difference this test could have caught: about 5 points. So this isn't "we couldn't tell" — it's a real answer. Market mood does not change whether a jump keeps going.

It does change how often jumps happen — slightly more per day during jumpy stretches. But once one shows up, the mood tells you nothing about what it'll do.

I built the tool specifically to find a flaw, and it reported back that there wasn't one. That's the most useful thing that happened all day.

04 / The Broken Rows

73 rows that were lying.

Free market data has bad rows in it, and bad rows do not announce themselves. In the earlier version of this study I found corrupted prices after writing the conclusions, and had to throw out an entire industry. So this time the cleaning came first.

Two kinds of nonsense turned up, and they were opposites.

Frozen prices

69 days where the stock's closing price was exactly the same as its opening price, to the penny. On a stock that supposedly jumped that morning, that's not calm — it's a stock nobody traded. Almost all of these had below-average volume. The price didn't hold steady; the price simply never updated.

Impossible days

4 days where a stock apparently doubled or tripled after opening, on 50 to 187 times its normal trading volume. Nothing like that happens. These are bookkeeping artifacts — when a company reshuffles its shares, the historical price series can carry the ratio through as one enormous fake move.

Every one of those 4 would have counted as a spectacular winner. All four sat in the same six weeks of spring 2023.

They add up to 2.3% of the data. I didn't delete them — they're filed with a note saying why each one went, so if a number ever shifts after cleaning, I can point to exactly what shifted it.

05 / What This Doesn't Tell You

Where to get out. Which is most of the trade.

"Kept going" here means one specific thing: the stock closed higher than it opened. That's a single strategy — buy at the open, sell at the bell, never flinch. It says nothing about how far it ran, when the best moment to leave was, or how much pain you'd sit through in between.

Measuring that needs minute-by-minute prices, and brokerages only keep about three weeks of those. You can't go back and get 2023. So that half of the question isn't answered here, and no amount of cleverness with the data I have will answer it. It has to be collected going forward, one day at a time.

Two smaller cautions. These stocks close lower than they open more often than not, on ordinary days too — so the numbers are a gradient to read against each industry's own normal, not a promise. And the AI group's quiet-day floor has been drifting upward year over year, meaning that end of the pattern is fading in that industry even as biotech's holds steady.

06 / Open Threads

What would make this real

  • Collect the exitsSave the minute-by-minute prices on live jump days as they occur. Thirty or forty is enough to start asking where to sell, instead of assuming you hold to the close.
  • Prune the universeA handful of companies produced most of the frozen-price rows. If a stock can't reliably be traded, an alert on it isn't worth receiving.
  • Test a different moodMarket mood was measured on a broad biotech fund. A crowd-behavior signal might respond to something more specific — how many small stocks are jumping on the same day, for instance.
Glossary

Every term I used, in plain words

Gap
The distance between where a stock closed yesterday and where it opened today. Usually caused by overnight news.
Relative volume
How busy a day was compared to that stock's own recent average. Twice as many shares changing hands as usual is "2x."
Confidence interval
The wiggle room around a measurement. Fewer observations means a wider range and less certainty.
Hidden Markov model
A method that finds hidden stretches in data — calm periods and wild periods — without being told where they are.
Lookahead
Accidentally using tomorrow's information to explain today. The most common way a market study fools its author.
Statistical power
How big an effect a test could have detected. Without it, "we found nothing" and "we couldn't have seen anything" look identical.