Three Weeks, Three Markets: Two Green, One Red, All On the Record

Three weeks is a strange unit of market time. Long enough for a full mood cycle — greed, doubt, small panic, relief — short enough that most people compress it into "the market was fine, right?" It was not fine. It was three different markets wearing the same index.

Here is what actually happened between July 6 and July 24, according to our data, our forecasts, and our own scorecard — including the week we'd rather not talk about, which is exactly why we will.

The shape of it

The S&P spent the stretch going nowhere slowly: roughly -1.7% over the three weeks, with almost all of the damage concentrated in two ugly sessions. Under that flat surface, the dispersion was violent. Our daily AI Top Picks book — every pick logged at the open, marked against SPY, misses included — ran three very different weeks:

• Week of Jul 6: +2.4% compounded, 52% win rate across 102 picks. Best: ENOV +14.1%.
• Week of Jul 13: +4.6% compounded, 65% win rate across 74 picks. The easy week. They're never labeled in advance.
• Week of Jul 20: -3.5% compounded, 37% win rate across 43 picks. The bill.

Since launch, the compounded book sits at +12.4% against SPY's +1.26% over the same 23 trading days (snapshot dated 2026-07-26 — the full daily table, drawdowns and all, lives on the public track record).

Anatomy of an air pocket

The week of July 20 deserves the autopsy, because it's where the regime showed its teeth. Monday and Tuesday were a slow bleed (-1.6%, then -2.15% — our worst day since launch, with a 10% win rate on Jul 22 that we are printing here on purpose). Growth names that had carried the melt-up got repriced first and fastest: PTC -6.5% in a session, TEM -8.7% on the week, quality software names giving back a month of gains in days.

Then Friday happened: +1.88% with a 90% win rate, led by TMUS +5.1%. Same model, same process, opposite tape. That's not the model "figuring it out" — that's what a dispersion regime does. It pays patience in lumps and punishes leverage in between.

The lesson we keep re-learning: win rate is a weather report, not a verdict. Two of the three weeks were green, the red one was survivable, and the compounding math is doing what compounding math does.

The confounder we owe you: our own data went stale

Full disclosure, because a track record without incident reports is marketing: from the evening of July 21 through July 23, our time-series store hit a query-overload failure mode. Heavyweight scans were jamming its executor, cheap reads piled up behind them (we counted nearly 200 stuck queries at the worst), and parts of the platform were reading data that was staler than our freshness bar allows. The picks made on the morning of July 22 — our worst day since launch, the 10% win-rate one — were made inside that window.

Can we attribute the damage cleanly? No, and we won't pretend to: Monday and Tuesday's losses landed BEFORE the incident began, the whole tape was getting repriced that week, and a dispersion regime doesn't need our help to produce a -2% day. But stale inputs are a real confounder on the 22nd and 23rd, and you deserve to know it exists before you read our red week as purely a model result.

What changed: the store now runs hard concurrency caps with a bounded queue (overload fast-fails instead of thrashing), a watchdog restarts it if queries ever pile up again, and the heaviest scans were moved to precomputed snapshots so they can't jam the live path at all. Friday's +1.88%, 90%-win-rate session ran on the fixed pipeline. We'd rather show you the autopsy than hope you don't notice the scar.

What the rotation map says

Our sector rotation scan spent the stretch quietly rotating toward the unloved. The strongest rotation score right now belongs to the beaten-down consumer corner — cannabis names (ACB, SNDL, CGC) showing forward-forecast strength the recent tape hasn't priced yet. Digital health is a genuine coin flip: the model is split almost exactly down the middle (5 bullish, 6 bearish) across DOCS, HIMS, PHR and friends — which usually means the sector is about to pick a direction and pretend it was obvious.

The short-interest freak show

The fundamentals-change detector logged some of the wildest short-float prints we've seen: YYGH's reported short float went from 76% to a frankly cartoonish 525% of float. VEEE went from 1.5% to 54% in one snapshot. And on the other side, UPC unwound from 72.6% short to under 5% — someone got out, all at once. We treat single prints like these as reported-data events, not gospel — but clusters of them are a regime tell: crowded shorts are being both built and detonated at the same time. That is not a calm market's behavior.

The crypto undercurrent

While equities chopped, crypto kept its own schedule. DEXE ripped +80% in a day late in the stretch — our research desk had a full brief on it published within the hour of the move getting loud. And this very weekend, while stock traders touch grass, the alt board is running hot: EUL +55%, FORTH +40%, SHIB +27% in 24 hours.

That last part matters more than it used to: we expanded coverage this week to 404 cryptocurrencies — every liquid USD pair on our data coverage, with AI forecasts on all of them and live pricing on the screener. Weekend moves no longer happen off our radar.

What the research desk kept finding

Behind the scenes, the desk ran 114 deep-research briefs and completed roughly 3,300 AI reports over the three weeks. Reading across them, two themes kept recurring:

1. Guidance season is doing the sorting. The regulatory-filing tape was dominated by earnings and guidance events — and the market's reaction function was asymmetric all month: in-line guidance got sold, small beats got ignored, cuts got obliterated. When the punishment is that lopsided, the market is telling you what it's actually worried about.
2. The AI-adjacent premium is being re-audited, name by name. Multiple briefs — ADI, TDY, EH among the recent ones — landed on versions of the same question: is this an AI beneficiary or an AI story? The stretch's biggest single-name drawdowns came from names where the answer was "story."

All published briefs are free to read on the research hub, and each one links the live forecast for its symbol so you can check our homework in real time.

Positioning into the next stretch

We don't do predictions in prose — that's what the forecast pages are for, where they can be scored. But the setup into late July is legible: a flat index hiding violent dispersion, defensives and left-for-dead corners quietly bidding, crowded shorts detonating in both directions, and a crypto tape that refuses to sleep.

Regimes like this don't reward being loud. They reward being early, small, and honest about the misses. Two of those three are model problems. The third one is a culture, and it's the one this post is for.

---

Every number above is measured, dated, and auditable — the track record updates daily, misses included. Nothing here is investment advice; it's a lab notebook with the lights on.