The number that said "strong consensus" while everyone was selling
This week's Brief is an apology and a correction. Last Sunday one of our own headline numbers was wrong on three pages of this site, and it was wrong in the direction that flatters us — it made a stock look more loved than it was. I found it because the top of one page didn't match the table underneath it. Here is what broke, what it now says instead, and what else fell out of looking.
1. Taiwan Semiconductor was not the most-agreed-upon stock in anything
On the Innovation Edge page, TSMC sat at the top with 85% smart-money confidence and the label "Strong consensus". Underneath it, the same page listed the funds that hold it: nine of the fourteen we track in that strategy, none of which had added to the position, and seven of which had trimmed it. Those two things cannot both be true. The cause was dull and worth stating plainly: the confidence percentage was being calculated across all fifty investors we follow, then printed on a page about fourteen of them. TSMC is held broadly across the full fifty. Inside Innovation Edge it is held by nine funds that were mostly heading for the exit. The page was answering a question it wasn't asking. It now reads 52%, "Moderate — mixed signals", and sits seventh of ten rather than first. Every holding on every strategy page was overstated the same way — thirty of them, twenty-three by fifteen points or more, and never once in the other direction.
2. "Company insiders are buying" was a staff share scheme
Fund ownership alone can only carry a stock to 75% here. To be called a strong consensus it needs corroboration from Congress or from company insiders. TSMC had no congressional trades, so the last ten points came entirely from insider buying. Those purchases are real, public, and filed on SEC Form 4s — twenty-one of them by twenty people, thirteen vice presidents, six SVPs, an EVP and the chairman, between nine and 186 shares apiece, all on the same day in April, $80,497 in total across the lot. TSMC's own filing explains what they were: shares bought by the administrator of the company's Employee Stock Purchase Plan "pursuant to terms predetermined by the issuer". Nobody looked at the share price and decided anything. The site now reads those footnotes and ignores trades the filer had no discretion over. That cuts both ways, which is the point: of eight recent NVIDIA insider filings I checked by hand, four are sales made under a plan agreed months in advance, and those stop counting as a bearish signal too. They were never a signal either. There is also a floor now: eighty thousand dollars spread across twenty people is not insider conviction at a company of that size, and the honest output is no signal rather than a direction.
3. And six of the fifty investors weren't actually there
Checking the first two fixes turned up a third problem, and it is the one that bothers me most. The model was being scored over forty-four of the fifty investors it claims to track. Two of them had the wrong identifier: Dodge & Cox pointed at the mutual-fund company, which files a different form and has never filed a 13F at all, and Maverick Capital pointed at Lee Ainslie the individual rather than the firm he runs. The other four — Renaissance Technologies, Trian, ValueAct and Altimeter — file perfectly readable 13Fs that we were reading wrongly, opening the cover sheet instead of the table of holdings behind it. Renaissance alone is 3,184 positions that were silently missing every week. The fix is merged and takes effect with the next data refresh, which will move some of these numbers again. I would rather say that now than have you notice it yourself next Sunday.
4. What changed on the model
Genuinely very little, and after the week I've had I am not going to dress it up. The highest-conviction names are still Microsoft, Amazon and Alphabet — the stocks the largest number of tracked managers hold at the same time. The scoreboard is unchanged: in the backtest the model beat the S&P 500 in 12 of 16 quarters, which means it lost in four. Everything above is a correction to how the site describes what the filings say, not a change in what they say.