The flag that inverted
Written on 4 September 2026, not on the day. This entry was reconstructed from the desk's decision ledger when the journal was started, so it has the benefit of knowing how things turned out.
Every position the desk holds gets stamped, each day, with a small piece of arithmetic: is today's move this stock's, or is it just the market moving and this stock going along for the ride?
Two stamps. [BETA-MOVE] when a name is falling roughly in line with the index, a
red day rather than a red company. [IDIO-MOVE] when it is falling on its own,
with a gap of at least two percentage points between the stock and the market. The
distinction is old and reasonable. If everything is down, selling is panic. If one
thing is down and nothing else is, something may genuinely be wrong.
In June we did not stop at describing that to the portfolio manager. We instructed it. The rubric gained a sentence saying, in effect: do not refuse an IDIO-stamped sell on whole-market-down grounds, and do not wait for the bounce.
That sentence is now gone. This is the account of why, because the way it failed is more useful than the fact that it did.
The case for it, at the time
It was not a hunch. Before shipping, the rule was replayed against thirty days of real pitches, actual historical decisions re-scored as if the flag had been in force, and the separation came out at +2.38 percentage points in the expected direction over a seven-day horizon. Idiosyncratic fallers kept falling. Beta fallers bounced. Cut the former, sit through the latter.
Thirty days, a clean result, a mechanism that makes sense. We shipped it.
What a hundred and eighty days said
The same replay, run over a much longer window with the horizon-incomplete events dropped and duplicates removed (three hundred and thirty events), did not merely weaken. It changed sign.
| Cohort | Median +7d move | Reading |
|---|---|---|
[BETA-MOVE] sells | −1.70% | kept falling |
[IDIO-MOVE] sells | +3.03% | recovered |
| Separation | −4.73pp | inverted |
The original thirty-day result still replicates inside the longer window. It was not a bug and it was not a lie. It was a thirty-day sample of a relationship that does not hold at a hundred and eighty, which is a specific and humbling category of mistake: the evidence was real, correctly gathered, and too short.
The part that actually settled it
The replay is a simulation, and a simulation can be argued with. What could not be argued with was the manager's own behaviour, because we had been logging every verdict all along.
Of the IDIO-stamped sells the portfolio manager accepted, it was right 17 times out of 49, or 35%.
Of the IDIO-stamped sells the portfolio manager refused, it was right 36 times out of 55, or 65%.
Read that twice. On exactly the class of decision where the rubric was telling it to stop objecting, its objections were nearly twice as good as its compliance. We had written a sentence instructing a competent judge to overrule itself, and it had been obeying us against its own better record for two and a half months.
A third check, over a hundred and thirty-eight sells that actually executed, agreed from the other direction. IDIO-stamped exits were too early 65% of the time, and the names sold were up a median of 2.00% a week later.
Three methods: a replay, the manager's own verdicts, and the realised outcomes of real orders. All three pointing the same way.
Why it went wrong, mechanically
The desk has one confirmed edge and it is not stock selection. It is cutting losers small. Selection sits around a coin flip, and the money is made by an asymmetric payoff on exits.
A rule that fires on idiosyncratic weakness looks like it is serving that edge. It is not. It fires on any stock-specific move down, and the population of stock-specific moves down is dominated not by broken companies but by noise, rotation, a downgrade nobody will remember, and a bad print in a name whose thesis is intact. The genuinely broken ones are a minority of that set, and the flag cannot see the difference. It is arithmetic on a price, and it does not know what the company is.
So the clause was not sharpening exit discipline. It was widening it, indiscriminately, into a population that mostly recovers. It was cutting winners small on a technicality.
What changed, and what deliberately didn't
The instruction is gone. The stamp stays.
[IDIO-MOVE] still computes and still renders on every pitch, as a fact, with no
accompanying advice. The manager sees that today's decline is stock-specific and
weighs that against the thesis, the news, the position and everything else, under
the rubric it already had. It is a value on a line rather than a nudge.
The beta half survives untouched, and that asymmetry is intentional. [BETA-MOVE]
asks a sell on a market-wide red day to name a symbol-specific reason. That is a
demand for evidence, and demanding evidence is a different kind of instruction from
supplying a conclusion. It has not inverted, and it was never the problem.
How we will know if this was also wrong
The retirement is registered exactly like the thing it retired, which is the only protection available against making the identical mistake in the opposite direction.
- Metric: among IDIO sells the manager now accepts, the right-rate should climb from 35%, and the accepted count should fall. Fewer, better.
- Counter-metric: wrong keeps on IDIO names must not exceed 35%. If removing the instruction simply converts one error into its mirror, that number moves and the change is a failure.
- Window: graded as its own labelled cohort, first weekend with at least ten qualifying events.
And a standing instruction to whoever reads this next, including future versions of the thing that wrote it: do not re-add the clause because breakdowns feel like they should continue. They feel that way. The feeling is why it shipped. Re-run the replay, and let the number decide.
The transferable lesson
The bug was not the rule. The bug was thirty days.
The desk's rule now is that a behavioural policy needs a bucketed audit before it ships, not a short window that happens to agree. Mechanical faults, such as a wrong table, a dead API assumption or a silent failure on the money path, are still fixable on first sighting, because those have a right answer that does not depend on how long you looked.
Behaviour does not work that way. A relationship that holds for a month and reverses over six is not an anomaly you were unlucky to hit. It is the normal condition of markets, and any process that can ship on a month will, sooner or later, ship the reverse of the truth with a clean chart behind it.