JT
MenuClose

Research Lab · simulation only

Most of my trading and sports ideas fail. That’s the point.

Two small labs, one rule: every idea is written down with what would prove it wrong, tested on past data, and kept away from real money.

One lab tests stock-options ideas, the other sports markets. By my rough estimate, they need about half an hour of my attention a day.

See how an idea is tested ↓ See the evidence ↓
In one picture: ideas are tested on paper first. Most are crossed out, and real money stays locked away.
The result 1,282 ideas written down; 11 passed every check. All 11 are trading ideas. No sports idea is still standing, and no real money has been used.
On this page · 5 sections · 6 min read
The problem

Finding something interesting is easy. Knowing whether it’s real is the job.

A test on past data can draw a beautiful line that means nothing, especially when nobody wrote the question down first. That goes for a stock chart and for a sports hunch, like a team scoring more goals after a long midweek match.

Both labs make challenging the result part of the routine: a limited list of questions, fair comparisons, simulated money only, and a record of everything that got rejected. It’s the discipline a good A/B test needs too: the success rule is written before the data comes in.

In plain terms

Every idea has to survive the same loop.

01Write the question. Say exactly what the idea should explain.
02Decide what would prove it wrong. Set the comparison and failure rule before any data is checked.
03Replay history. Test against past market or game data and keep the records.
04Challenge the result. Compare it with a simpler explanation, simulated money only.
05Record what happened. Pass, fail, no real pattern, or too little data, all dated.
06Check before the next run. See what changed, and keep real money out.
Lab one · stock options

Most trading ideas do not survive a fair comparison.

  1. 01 · Tested446ideas tested and written into the record
  2. 02 · Still plausible38were not negative or incomplete after the first cut
  3. 03 · Passed11passed every check
408 stopped earlyAnother 27 looked better or repeated successfully but still did not qualify as a clean pass.
Passing is a high bar on purpose. 91.5% came back negative or incomplete. An idea can look interesting, beat one comparison, or repeat once and still fail the full set of checks.

Status as of September 3, 2026 The decision in the dated record: STAY ON PAPER. Nothing here has been sized for real money.

See every recorded result and the scale behind it
Every outcome stays visible, drawn against all 446 ideas: no real pattern, no usable data, failed, weaker, better, repeated, or passed. Blue marks the ideas that passed.
Each box holds the next. Volume shows effort, not that an idea works; the checks still decide what counts.
What one verdict record looks like, with the details redacted
The idea

A ████ options strategy can run without an unacceptable loss.

Written down

August 27, 2026, in the plan, before any result was judged against it.

What would prove it wrong

Its worst drop from a peak is more than 15% of the account.

Result

0 of 11 candidates passed. The best one’s worst drop was more than twice the account (214%).

Decision

Rejected. The record stays, with its date.

Every idea leaves a record like this, pass or fail. Strategy details stay private; the dates, the rule, and the verdict don’t.
Lab two · sports markets

Every sports idea gets a written “prove me wrong” rule before it’s tested.

It’s tempting to notice a pattern and keep adjusting the test until the pattern survives. I flip that: before any test runs, I write down the condition that would end the idea. Researchers call that a falsifier.

The automation refreshes the data; I write the rules, read the results, and decide whether an idea moves on. The answer is allowed to be boring.

See a written “prove me wrong” rule (illustrative)
Illustrative example, not a real idea
The idea

After a team plays a long midweek match, its next weekend game produces more goals than usual.

What would prove me wrong

The pattern disappears once opponent strength and home advantage are accounted for, or fails in a season kept out of the first test.

Time limit

500 games, or the end of the season, whichever comes first.

If it fails

Rejected, with no second chance.

The rule exists before there is any reason to want the idea to survive. This is what a written rule looks like before a question enters the system.
  1. 01 · Written down836questions entered the research record
  2. 02 · Ever looked promising3survived enough checks to be called promising for a while
  3. 03 · Active snapshot0remained active on October 14, 2026
517 stopped outrightAll three promising ideas were later rejected. Two were disproved by my own follow-up checks: they were the same evidence counted twice, and their apparent win came from duplicated records and prices no one could have traded at. The third didn’t hold up on data kept out of the first test.
The steep drop is the point. Across 972 reviews of the data, 65 questions were replaced by a better version and 26 were retested on fresh data. 41,518 simulated positions are logged, and none used real money.
See the detailed rejection reasons
Only three questions ever looked promising. Each bar is drawn against all 836 questions, with labels and values on every row. Blue marks the promising group; rust marks questions that failed a fair comparison or looked suspicious.
System checks

The research is only as trustworthy as the system running it.

The steps that produce a conclusion get checked on a schedule, not only when a result looks surprising.

Tests 1,107 automated tests pass in the trading lab. They exercise 48.8% of the code: a safety net, not full coverage.
Problems caught 12 problems logged; the typical one was noticed within about a day. Each stays in the log with what broke and when it was caught.
One example A broken run was kept out of the results. A health check stopped an incomplete run before any conclusion was recorded.
Automation 75 scheduled tasks keep both labs running. I read results instead of watching every step; what the system may do stays the same.
What it took

The interesting part is how often the answer was no.

How long
Built mostly between early July and mid-September 2026, going by the project’s file dates. Both labs keep running on their own schedules.
What broke
An incomplete trading run, repeated after the fix and kept in the log. On the sports side, two ideas that looked promising were later disproved by my own follow-up checks.
Do differently
Show the untested gaps sooner, and show why each idea stopped at the check that stopped it, so nobody has to guess.
Evidence record

Every number on this page, with its source.

Open the full evidence record

This public record contains counts and process measures only. Strategy names, which stocks, leagues, or markets were tested, cutoffs, position sizes, prices, profit figures, and anything that could reveal a winning idea are deliberately left out. The checked summary stays public; the larger working record remains private.

One row is one recorded verdict or question, not one trade or match. An idea can be tested many times, and it stays in the record after it is rejected. Nothing here was sized for real money.

Trading lab (stock options)

The claim

446 ideas tested and recorded
11 passed every check
408 negative, noise, or not enough data
91.5% came back negative or incomplete
STAY ON PAPER standing decision

The checks

1,107 automated tests passing
48.8 % of code exercised by tests
12 problems logged
24.0 typical hours until a problem was noticed
226 planned combinations tested
583 planned combinations in total

The scale

26,088 historical tests
1,869,008 simulated trades
44 scheduled tasks
September 3, 2026 status record generated (03:01)

Sports lab

The claim

836 questions logged
3 ever looked promising
0 promising ideas active at snapshot
517 rejected total
2 later disproved

The checks

100% written-first rate
100% rejection-rule rate
717 timing-scramble checks
33 deeper comparisons
26 fresh-data reruns
65 questions replaced
972 total data reviews

The scale

3,207,138 data points
356 batches
15 question types
6 market groups
41,518 simulated positions
37,606 simulations settled
3,912 simulations still open
31 scheduled tasks
0.5 hours old, newest data at snapshot
October 14, 2026 snapshot taken (11:30 UTC)

How all 836 sports questions reconcile

The first 517 break down into 393 process failures, 118 data failures, 4 grouping problems, and 2 suspicious results. The other 319 reached later checks: 126 failed, 71 did not have enough evidence, 45 failed a fair-comparison check, 65 were replaced by a better version of the question, 9 had no clear answer, and 3 looked promising. Together those groups account for all 836 questions.

Talk shop

Seen a backtest that looked too good to be true?

I like comparing notes on testing ideas fairly, whether it’s a trading rule, a sports hunch, or a marketing report.

Connect on LinkedIn →
Keep going
Next project Boulevard 21 A blackjack card game, tested by simulation. See the game → Or browse All five projects Marketing work and things I built for fun. See the projects →