Pre-Earnings Read · both report today · Aug 4 2026
Mosaic & Toast, before the print
Not a fundamentals recap — the setup: what a company's own earnings history, its
post-report drift pattern, and its insiders were signaling before today's calls. Assembled from the
entity graph + filings, point-in-time.
🔀
The naïve glance gets both backwards. MOS looks ugly (a loss, collapsing margins);
TOST looks beautiful (83% beat rate, expanding margins). But the setups flip the nuance — MOS is a cyclical-bounce
wildcard, and TOST is a textbook
beat-and-fade.
MOS
Mosaic · fertilizer / ag (cyclical)
Beats own trend33%
Last 2 quartersboth missed
Latest op margin−12.4%
Drift after a miss−120 bps
Drift after a beat+171 bps
Insider signalnone
- Deep in the cyclical trough — gross margin 7.9%, an EPS loss
- Its own drift base-rate punishes misses; the last one drifted −9 to −12%
- The offset: cyclical names bounce off a low base — that's the wildcard
The call
Cautious → fade lean
confidence low 0.50 · cyclical wildcard
TOST
Toast · restaurant POS / fintech (growth)
Beats own trend83%
Revenue growth+22% YoY
Op margin trendexpanding 3→7%
Drift after a beat−133 bps
Surprise trenddecelerating
Insider signal5/5 C-suite selling
- Genuinely excellent business — consistent beats, margins inflecting up
- But priced for perfection: the stock drifts down even after beats
- Every C-suite officer sold into the print (CEO, CFO, President, +2) — zero buyers
The call
Beat & fade
confidence moderate 0.62 · 3 tells align
Why these reads have teeth (and why we're logging them)
The load-bearing signal — post-earnings drift conditioned on the surprise — isn't a hunch;
it's a PIT-validated market anomaly (+106 bps, t-stat 5.9). Layer on insider breadth (all five Toast
officers selling is a stronger tell than the dollar amount), and you have a real feature set. Both calls above
are falsifiable and logged — direction + confidence, stamped as-of today — so when the prints
land tonight we score them against the actual reaction. Do that for every name before every event and the log
becomes the training corpus for the prediction model. This is entry #1 and #2.