Good at funds, bad at stocks

2026-09-28 weekend-reviewmeasurementselection

The desk lost ground in both weeks of the fortnight. The S&P rose in both.

At the last review, a fortnight ago, the desk was 4.14 points ahead of the index over six weeks. By this Saturday it was 1.08 points behind.

A fortnight is noise, and we know it. So we asked a narrower question with a much larger sample: when the manager (the model that accepts or refuses each pitch) says yes to a buy, does that buy do better than the ones it said no to?

The question, asked properly

We took every buy pitch since 1 May. That is 1,152 pitches, reduced to 735 once the same symbol pitched twice on the same day is counted once. Each was scored against its own market, so US names against the S&P and London lines against a London-listed S&P tracker, over the next one to twenty sessions.

The answer depends on what kind of thing is being bought.

London lines on this desk are almost all sector funds, about 88%. US lines are almost all single companies, about 97%. On new London entries the manager's picks beat its own rejects at every horizon we measured:

That is n=37 accepted against 143 rejected at five sessions, and the gap holds in both halves of the window.

On US single stocks, over the full window, the accepted buys did 1.97 points worse than the rejected ones at five sessions. Since August the two groups are indistinguishable (+0.96% against +0.88%, n=21 and 50).

On single stocks, the manager's scores told us nothing about what would happen next.

Adds to a fund we already held showed no edge either. Whatever the manager is good at, it shows up when choosing a new sector and disappears when topping one up.

What was steering it

A model repeats what its instructions tell it with confidence. So we read the instructions for confident claims the data did not back, and found three.

The first was a line saying analyst upgrades were the desk's measured edge, worth +2.99%. Measured again, accepted analyst-action buys went −0.71% at five sessions (n=52) while the ones it refused went +2.09% (n=75). The manager had been quoting the +2.99% back to us in live verdicts. It was citing our own brochure.

The second gave a buy extra credit for timing when the stock was already running on its own news. Those buys went −1.39% at five sessions (n=54). The refused ones went +1.51%.

The third showed the manager each analyst's "track record" beside an instruction to weight by accuracy. The track record was a trade count. It had no accuracy in it.

All three were deleted. No rule was added in their place.

We also relaxed one veto. A buy far below its 200-day average was forced to the lowest timing score, and the manager refused 36 of 39 such pitches. The pool it refused beat the rest by 4.3 points at three sessions.

Before shipping, every recorded verdict was re-scored under the new rules. The average accepted single-stock buy moved from −0.87% to +0.01% against the S&P at five sessions. Almost all of that came from May to July. Since August about four trades would flip, so the honest forecast is a small effect.

What we did not do

The obvious conclusion is to stop buying single stocks. Steve looked at the same table and kept them live.

That is a reasonable call on this evidence. Since August the single-stock picks are no worse than chance, and the confidence interval on n=21 is wide enough to hold a real edge or a real loss. So the decision is pre-registered instead: once 40 accepted single-stock buys have matured, if their average five-session return against the S&P is zero or below, the question comes back. The count stands at 21.

The early reading is not encouraging. This week's nine single-stock buys were −3.56% against the market by Friday. Four catalyst buys were sold inside 48 hours at a loss. In the same fortnight, five exits taken at the profit target written down at purchase all paid.

What paid this fortnight was the plan written before each trade. What cost money was judgement exercised afterwards.

Verdicts that went against us

Changes are graded on their own record. Three did badly.

• Moving the US morning sweep later was meant to stop it firing before US data had settled. The new slot landed on the same minute as the daily position review, so it lost the race for the run lock every day. It did not run once between 17 and 25 September. It now runs at 10:50 New York and retries once if blocked.
• A flag for stops that were already broken at purchase (the condition the bot wrote down that would prove itself wrong, already true on day one) fired nine times this week. All nine were misreadings of numbers that were not stops: a price target, a moving-average window, a date. The manager cited the false flag in eight verdicts. The scanner was fixed rather than removed, because the one broken stop it did catch this month, on 11 September, was genuine.
• The trend hunter, which writes predictions that nothing can trade on until they are graded, missed its pre-registered bar for the fifth time. Two of its ways of reading the news have never been switched on. Whether to narrow it or keep paying for it is an open question for Steve.

One retirement is still pending. On 2 September the desk stopped telling the manager not to refuse a sell just because the whole market was down, when the stock was falling faster than the market. Since then it has accepted five such sells and none was right. The cohort holds nine such events against a bar of ten, so that verdict waits.

Next Saturday: the first read of the three deleted claims, and how close the single-stock count has got to 40.