TAILFIFTY · RESEARCH

Will Valero (VLO) fall by at least 6% at any point within the next 30 days as Gulf Coast storm risk rises?

markets30dasked 2026-08-20 · resolves on VLO.close min_return<= -0.06

Committee19%

OutcomeNOBrier 0.036

Committee

Weather Analyst22%was 21%

Public Markets Analyst22%was 42%

Commodities Analyst20%was 22%

Quantitative Researcher20%was 22%

Market Positioning Analyst18%was 30%

Causal Skeptic15%was 19%

Top disagreement

D's storm-intensity/refinery-utilization proxy (d3) reportedly spikes from a baseline ~185 to 355-366 in mid-August 2026 (read as an active, unusually large current threat), while E's storm-risk index (d8) reportedly declines for six straight sessions from 2.913 to 1.167 through 2026-08-20 (read as the named catalyst fading, not rising, at the as-of date). This is the single most decision-relevant conflict because it bears directly on whether the question's own premise -- 'as Gulf Coast storm risk rises' -- is even true at t=0, and neither analyst reconciles it against a primary lead-time signal (NHC cone probability, days-to-landfall). I did not adjudicate which index is correct: the two series use incompatible, unlabeled units/scales (185-366 vs 1.2-2.9) and could plausibly be measuring different things (refinery utilization/outage proxy vs. a normalized storm-day count), so neither is verifiable as a contradiction of the other from what's in the record. I resolved this by treating it as a confidence-reducing unresolved fact rather than resolving in either direction, and by noting that it is largely moot for the point estimate: every storm-conditional statistical test in the panel (conditional lift n=17-49, four regression specifications, four event-study constructions) is null or mildly positive-signed regardless of which index reading is right, so the disagreement affects tail-risk sizing and confidence width, not the central estimate. I widened the credible interval's upper bound modestly (to 0.27) to carry this unresolved risk rather than compressing it into the point estimate.

Synthesis

SYNTHESIS Base rate: The panel converges, after retracting two flawed parametric (Brownian-barrier) estimates, on an empirical unconditional P(30-day path-min <= -6%) of 0.18-0.23, built from two genuinely distinct bootstrap constructions (A/C ~0.18, D/F ~0.23). This is the correct anchor for a touch-probability question and is well-supported. Causal evidence on the storm-conditioning hypothesis: every independent test design applied to the "storm risk rising -> VLO falls" causal chain returns null or mildly opposite-signed results -- conditional-lift tests are underpowered and sign-inconsistent (n=17-49), three OLS-HAC regressions on storm-region/wind-intensity variables find no explanatory power beyond RBOB/WTI, four event-study constructions find flat-to-positive (never negative) CARs after storm-adjacent dates, and two factor-decomposition regressions confirm VLO's dominant explainable driver (RBOB/XLE/SPY beta) is bullish for refiners during exactly the kind of disruption event the question describes. F's version of the event study is the most rigorous single test in the panel because it alone controls for the "most 30-day windows are up anyway" confound via a random-date bootstrap null, and it still finds a positive-signed, non-significant result. Given this, I do not adjust the consensus probability upward from the empirical base rate for the "storm risk is rising" premise -- the weight of independently-designed evidence argues there is no detectable incremental risk from that specific catalyst, and if anything a mild bullish tilt via the crack-spread channel. I do carry forward a small residual premium (landing at 0.19, near the low end of the empirical band rather than its midpoint) for the one channel nobody tested: an actual direct hit on Valero's own Port Arthur/Corpus Christi/St. Charles assets, which is not ruled out by any of the sector-level null results and represents genuine, unquantified tail risk consistent with the question's premise. Top disagreement (D's storm index spiking vs E's storm index declining at as_of) is resolved as unresolved: both indices are unlabeled, differently-scaled constructs that cannot be reconciled from the record, and critically, it doesn't change the point estimate because the storm-conditional statistical tests are null regardless of which reading is correct. I widened the credible interval's upper tail to carry this. TRADEABILITY: I do not consider this a tradeable directional short. The estimated probability (0.19) sits at or below the unconditional historical base rate, meaning the "rising storm risk" catalyst named in the question does not, on this evidence, justify incremental downside positioning in VLO specifically -- the more statistically robust findings (factor loadings, event-study CARs) argue for a neutral-to-mildly-bullish read on Gulf Coast disruption for a refiner. We also lack any options/IV/skew data, so we cannot assess whether markets already misprice (or correctly price) even the small residual tail risk, which is a real analytical gap, not an argument for or against a trade. If a portfolio wants exposure to this thesis at all, the only defensible expression is a small, cheap, convex tail hedge -- e.g., a far-OTM (12-18% below spot) 30-45 day VLO put or put spread, sized as pure insurance against the untested idiosyncratic direct-hit-on-Valero's-own-assets scenario, not as a high-conviction bet that "storm risk rising" itself predicts a >=6% drawdown. I would not size this above a token hedge allocation given the negative expected value implied by the null causal evidence, and I would not short outright or buy at-the-money puts. Monitoring the two conflicting storm-index constructs against actual NHC track data over the coming days is the single highest-value next step before any position sizing decision.

What is what

Observed

VLO closed at 341.5 on 2026-08-20, within 2.5% of the trailing series high of 350 (reached 2026-08-18) (Analyst E, d1).
The two most recent daily returns before as_of were -1.08% and -1.37%, the first back-to-back down days in the recent window (Analyst E, d2).
660-day VLO daily return series: mean 0.1456%/day, sd 2.233%/day (Analyst B, d2) -- one analyst reports sd=2.42%/day for what is claimed to be the same series (Analyst E), an unreconciled data discrepancy.
Unconditional empirical frequency of a 30-day path-minimum return <= -6% clusters at 0.18-0.23 across four differently-constructed samples (A n=631=0.1807, C=0.1806 [same series as A], D n=612=0.2304, F n=570=0.2316).
OLS-HAC regressions of VLO returns on storm-region dummies / wind-intensity proxies, controlling for RBOB/WTI, show no significant coefficients (p=0.40-0.993, R²=0.0007-0.008) across three independent specifications (A, C, D), all flagging severe residual autocorrelation (DW=0.07-0.09).
Event-study CARs after storm-adjacent events are flat-to-positive, never significantly negative, across four window/sample constructions (A: +1.48% p=0.907; D: +2.0% n.s.; E: +3.5% p=0.634; F: +0.57% p=0.974 vs random-date bootstrap).
VLO loads positively and significantly on market/crack-spread factors: RBOB +0.239 (p<1e-5) and XLE +1.10 (p<1e-68) per B's t2; RBOB 0.40 and SPY 0.64 (R²=0.28, p<1e-6) per F's t7 -- two independently-specified regressions agreeing that refiner co-movement is dominated by market/margin beta, not a storm-damage channel.
Two mutually-unreconciled storm-index readings exist at the as_of date: D's d3 proxy rising sharply (~185 to 355-366) vs E's d8 index declining for six sessions (2.913 to 1.167).

Inferred

The storm-conditional lift test (29.4% vs 18.1% base, 1.63x, n=17, p=0.209) is underpowered and not statistically distinguishable from the null; it is cited identically by A and C and should count as one piece of evidence, not two.
The market/crack-spread factor loadings (bullish RBOB/XLE/SPY beta) imply that a Gulf Coast disruption is more likely to transmit through refining margins (positive for VLO) than through direct equity selloff, undercutting a mechanical storm-risk-rising to VLO-falls causal chain.
B's original 0.42-0.51 Brownian-barrier estimate was a specification error, not a correction to the empirical base rate: the reflection-principle formula requires i.i.d. increments, which the data violate (DW~0.07-0.09), and the empirical base rate is already a path-minimum (not terminal-return) statistic, so no adjustment for path-dependence understatement was warranted at that magnitude.
D's n=15-trade backtest (Sharpe 1.83, net-long during active storm windows) is too fragile (self-flagged) to meaningfully shift the estimate in either direction and is treated as near-zero-weight evidence.

Modeled

Committee-constructed empirical base rate for P(30-day path-min <= -6%) via bootstrap resampling of historical VLO return windows, landing in a 0.18-0.23 band across four sample constructions.
OLS-HAC regressions of VLO daily returns on storm-region dummies/wind proxies controlling for RBOB/WTI (three independent specifications, all null).
Factor-decomposition regressions of VLO returns on RBOB, XLE, and SPY (two independent specifications, both showing strong positive/bullish loadings).
Event-study CARs following storm-adjacent event dates vs a random-date bootstrap null (four constructions, all flat-to-positive).

Predicted

Consensus probability that VLO touches a >=6% intraperiod drawdown at some point in the 30 days from 2026-08-20: 0.19, CI95 [0.13, 0.27].
Expected 30-day terminal return for VLO: modestly positive, roughly +2% to +3%, consistent with the unconditional daily drift (~0.145%/day x ~21 trading days) and the directionally positive (though statistically insignificant) post-storm-event CARs found across four independent event studies. This is NOT the same variable as the touch probability -- a mildly positive expected terminal return is fully consistent with a ~19% chance of touching -6% intraperiod given realized daily vol of ~2.2-2.4%/day.

Uncertain

Whether Gulf Coast storm risk was actually rising or fading at the 2026-08-20 as_of date -- D's proxy (d3) and E's index (d8) directly conflict and neither is reconciled against a primary source (NHC cone probability, days-to-landfall).
VLO's return-distribution behavior conditional on an actual hurricane-warning window targeting its own Port Arthur, Corpus Christi, or St. Charles assets specifically (as opposed to sector-wide storm-region dummies) -- no analyst ran this test; it is the single largest unresolved tail-risk gap, flagged independently by B, C, D, and F.
Whether current options-market implied volatility or skew already prices in elevated storm-driven crash risk for VLO -- no analyst had IV/skew data (E flags this explicitly as a real gap), so we cannot assess whether a short-vol or long-put trade would have positive expected value net of pricing.
The true magnitude of the continuous-monitoring vs. close-to-close touch-probability gap (D's point that daily-close sampling mechanically undercounts intraday touches) -- directionally raises the true touch probability above the empirical band's midpoint by an unquantified amount.
Whether the ~5pp spread in the base-rate anchor itself (18.07% vs 23.04% vs 23.16% across A/C vs D vs F) reflects genuine sample/window construction differences or a reconcilable discrepancy nobody resolved.

Falsification

VLO trades at or below 320.9 (341.5 x 0.94) at any point between 2026-08-20 and 2026-09-19: the predicted event has occurred, falsifying the 'no large drawdown' call.
An NHC-tracked storm system makes a direct hit (landfall within ~50 miles, sustained winds >=74mph) on Valero's Port Arthur, Corpus Christi, or St. Charles refinery complexes within the 30-day window: this would test the one genuinely unresolved tail-risk channel that no committee analysis covered, and a subsequent large VLO drawdown in that scenario would validate the idiosyncratic-landfall thesis rather than the sector-level null findings.
VLO realized volatility over the 30-day window materially exceeds the historical ~2.2-2.4%/day (e.g., >3.5%/day realized), which would indicate the empirical base-rate anchor itself understates current regime risk and require the estimate to be revised upward independent of the storm-causality question.

← all research

Every forecast is pre-registered with a resolution rule and scored when it resolves. Numbers come from deterministic tests over point-in-time data. Every seat on the committee is an AI model with a role prompt, not a person or a licensed professional. TailFifty is an educational research experiment run by an individual. Probabilities and paper trades are AI-generated and simulated; there are no real positions. Nothing here is investment, trading or betting advice, or an offer or solicitation. Not a registered investment adviser, broker or CTA. Hypothetical and simulated results do not reflect real trading, liquidity or costs, and past forecast accuracy does not guarantee future results. Data: NOAA, NHC, CFTC, Open-Meteo (CC BY 4.0), CoinGecko, Kalshi and others; derived figures only. © 2026 Harrison Johnson. About and disclosures.