VST Earnings Recap: Solid Print, Soft Reaction — The Contracted-MW Feature Wasn't There

The Pre-Print Cohort Read

We published VST's live cohort this morning: 17 AI-power-era analogs (CEG, TLN, post-2023 VST), 5-day forward median +1.2%, IQR [-5.4%, +9.8%]. The dominant feature was contracted-MW disclosure: new hyperscaler contract + raised long-term capacity guide → 5-day +8.7% (n=5, hit-rate 100%); no new MW disclosure + flat capacity guide → 5-day -4.6% (n=4, hit-rate 25%).

What Actually Printed

Q1 2026 Net Income of $1,029M (vs. a Q1 2025 loss), Ongoing Operations Adjusted EBITDA of $1,494M. Full-year 2026 EBITDA guide reaffirmed at $6.8-7.6B. Investment Grade credit upgrade from a second major rating agency. Solid fundamental print.

Reaction: -1.3% intraday — sitting in the cohort's 'no-new-contract' bucket. The print didn't include a new hyperscaler contract announcement or a raised long-term MW capacity guide. The 18-month historical pattern of VST post-earnings reactions averages -1.69% — almost exactly where today's print landed.

What The Cohort Got Right

The feature attribution called this one cleanly. The bullish-bucket scenario required two things: a new contract announcement AND a raised capacity guide. The print delivered neither — it was a 'reaffirm and execute' print, not a 'we just signed a new hyperscaler' print. The historical analog to this scenario is the cohort's lower bucket, and the realized -1.3% reaction sits inside the bucket's expected range.

Conformal 80% band was [-7.9%, +12.3%] for 5-day forward. Today's intraday move (-1.3%) is well inside the lower IQR bound (-5.4%) and the conformal lower bound. Successful calibration — the cohort told you what the bear case looked like, and the print delivered the bear case.

• Pre-print 5d cohort median: +1.2%; bear-bucket median: -4.6%
• Conformal 80% band: [-7.9%, +12.3%]
• Actual reaction: -1.3% intraday
• Verdict: cohort feature attribution called the bear scenario; print delivered it

The Calibration Lesson

VST's print is the canonical 'cohort works as designed' outcome. The feature attribution gave you a binary read: contracted-MW announcement = bull, no announcement = bear. The print resolved the binary. The reaction landed where the historical analog set predicted.

For agent systems, the takeaway is that strong feature attributions are tradeable — when the cohort surfaces a feature with high importance and a clean bucket separation (here: ~13pp delta between bull and bear buckets), the trade is to wait for the feature to resolve and then size the position to the bucket the print lands in. VST's reaction was small in absolute magnitude, but it was directionally clean and inside the conformal band.