NVDA Pre-Earnings Drift: A Wide Cohort, A Time-Decay Feature

NVDA's Pre-Print Cohort Is Wide By Design

Nvidia's earnings reactions are among the highest-magnitude in mega-cap tech, with 5-year median |move| of ~9%. The pre-print cohort, however, is structurally wide — the cohort_analyze pull on 2026-05-06 retrieved 258 historical analogs with a tightness score of 0.24. That's expected: NVDA's forward distribution is dominated by the print itself (datacenter revenue, gross margin, hyperscaler color), not by the pre-print drift.

The 5-day forward median is +0.09% with IQR [-4.7%, +3.8%], hit-rate 51%. The 10-day median is +0.25%, IQR [-6.8%, +9.2%]. Wide and symmetric — the pre-print shape isn't telling us much yet.

Time-Decay As A Feature

The top 5-day feature was days_since_earnings with positive direction (importance 1.11). That's the decay signal — analogs further from a recent earnings event had stronger forward returns. Read carefully: this means the closer the cohort match was to its own next earnings date, the weaker the forward returns. NVDA is currently ~2 weeks out from its print, which is right at the boundary of this feature's effect — slightly biased toward the cohort's lower bound.

This is a classic 'wait for the print' setup. The pre-print cohort gives you a base rate but no actionable directional signal.

• n=258, 5d median +0.09%, IQR [-4.7%, +3.8%], hit-rate 51%
• 10d median +0.25%, IQR [-6.8%, +9.2%]
• Cohort tightness 0.24 — wide, not actionable pre-print
• Top 5d feature: days_since_earnings (positive direction, importance 1.11)

What The Print Will Re-Anchor

Once NVDA prints, the post-call cohort will tighten to ~0.55-0.65 based on prior post-print retrievals. The features that historically separate the upper IQR from the lower IQR are: datacenter revenue YoY growth rate, gross margin trajectory, and hyperscaler capex commentary mix (the 'big four' guidance for the next two quarters). Hits on all three map to the +9% upper band of the 10-day IQR; misses on margin map to the -7% lower band.

Agent systems should treat NVDA's pre-print cohort as a 'wait' signal. The print is the trade.