Test of GEX/DEX/VEX/CHEX on 1,972 SPY days: raw GEX looks great, dies after VIX + ATM IV controls
**TL;DR**: I tested the four dealer-exposure Greeks (GEX, DEX, VEX, CHEX) against next-day SPY outcomes on 1,972 end-of-day snapshots from Apr 2018 to Apr 2026. Raw GEX looks strong (Spearman ρ = -0.36, p ≈ 10⁻⁶⁰). After controlling for VIX and ATM IV, it drops to ρ = -0.03 (not significant). DEX has no signal even raw. VEX is basically a VIX proxy. CHEX is borderline noise at EOD. Hypotheses were pre-registered before the stats.
**What I did**
For each of 1,972 SPY trading days I joined same-session GEX, DEX, VEX, CHEX, gamma flip, VIX, ATM IV - with next-day return, next-day realized vol, next-day IV change.
* Sort days into 5 equal groups ("quintiles") by the exposure I'm testing, check the next-day outcome.
* Then residualize: regress the signal and the outcome on VIX (or VIX + ATM IV), take the residuals, re-run. If the signal has independent information, it still works. If it was a VIX proxy, it dies.
* **I wrote the hypotheses down in a file before running any stats.** Saves me from convincing myself of whatever the data happened to show.
*Quick glossary for non-quants:*
* *Quintile = 5 equal groups. Q1 = lowest 20%, Q5 = highest 20%.*
* *Spearman ρ = rank correlation, from -1 to +1. +1 = ranks line up perfectly. -1 = perfectly opposite. 0 = random.*
**The raw GEX result - this is what sells subscriptions**
Next-day realized vol (%, annualized) by GEX quintile:
|GEX quintile|n|Mean next-day RV|
|:-|:-|:-|
|Q1 (most negative)|395|**17.0%**|
|Q2|394|**18.6%**|
|Q3 (neutral)|394|12.7%|
|Q4|394|9.2%|
|Q5 (most positive)|394|**6.3%**|
Spearman ρ = **-0.36**, p ≈ 10⁻⁶⁰. On an 8-year sample that's statistically real. The Q1/Q2 mean inversion is COVID-era outliers; medians are cleanly monotonic.
**The twist - what happens when you control for VIX and ATM IV**
|Signal → Outcome|Raw ρ|After VIX ctrl|After VIX + ATM IV|
|:-|:-|:-|:-|
|GEX → next-day RV|\-0.36|\-0.14|**-0.03** (p=0.18)|
|DEX → next-day return|\-0.03|\+0.01|\+0.02|
|VEX → next-day IV change|\-0.16|\-0.05|\-0.01|
|CHEX → next-day return|\-0.05|\-0.01|\-0.00|
GEX survives a VIX-only control (weakened). Dies when ATM IV joins. The other three never had much to lose.
**The real killer - double-sort heatmap**
5×5 grid: rows = VIX quintile (V1 calmest, V5 most stressed), columns = GEX quintile. Cell = mean next-day realized vol (%).
|VIX GEX|Q1|Q2|Q3|Q4|Q5|
|:-|:-|:-|:-|:-|:-|
|V1 (low)|8.0|7.3|6.6|5.2|5.1|
|V2|11.7|10.3|8.8|6.5|6.0|
|V3|12.0|12.1|12.2|9.6|8.6|
|V4|15.9|15.4|17.2|12.3|8.1|
|V5 (high)|20.6|24.9|**37.7**|21.7|15.9|
Rows V1–V2 look textbook. V3–V4 are close but not strictly monotonic. **V5 - the stressed regime where you actually want a signal - is a non-monotonic mess.** Middle GEX has the highest RV in the entire grid.
The GEX regime split confirms it: on the top-VIX-quartile subset of 493 days, the top-vs-bottom GEX next-day RV difference is **-1.89 vol points, t = -0.78, p = 0.44**. No signal at all in the regime people care about most.
**One-line verdicts**
* **GEX** \- useful regime descriptor (positive = dealers absorb flow = pinning; negative = dealers amplify = wider tape), but no independent alpha over VIX + IV.
* **DEX** \- zero predictive content. Top-vs-bottom next-day return diff = 0.00%, p = 0.97.
* **VEX** \- 72% correlated with VIX, 76% with ATM IV. Strip those out, nothing remains.
* **CHEX** \- 54.9% sign-agreement with next-day return raw (p = 10⁻⁵). Dies under any rank control. **EOD only**\-this study does not test the intraday last-hour-charm narrative; minute-level data is needed for that.
Collinearity matters too: DEX and VEX correlate with each other at **-0.89**. They are two sides of one coin.
**What I'd actually do with this**
1. **Ignore GEX in high-VIX regimes.** In the top VIX quartile, the signal is noise.
2. **In calm regimes, GEX as a vol-label is fine** \- just know VIX would have labeled the regime the same way, and VIX is free to quote.
3. **Stop counting GEX/DEX/VEX/CHEX as four independent signals.** They span roughly 1.5 effective dimensions, and VIX+IV already covers most of those.
4. **If you want real independent edge, the signal has to be orthogonal to VIX and IV.** None of the four dealer-exposure Greeks are. Candidates worth running the same test on: VIX term structure (VIX/VIX3M), realized-minus-implied skew, order-flow imbalance, VRP residuals.
**Honest caveats** (these matter more than the headline)
* EOD only. No intraday CHEX test.
* SPY only. Single names will likely look worse (less dealer hedging, more idiosyncratic drift).
* Linear OLS residualization - a nonlinear model could extract edge OLS misses.
* Correlation, not PnL. A residual ρ of -0.14 does not automatically become a profitable strategy after costs, slippage, and execution lag.
* 2022 is the only real bear year. High-VIX regime inference rests on \~493 days.
* Dealer-sign convention: positive = dealers net long that Greek. Vendors using the opposite sign will see everything flipped - conclusions identical.
**Data and full write-up**
Full article with regime splits, train/test (70/30), correlation matrix, per-exposure verdicts, limitations, and downloadable CSV artifacts: [https://flashalpha.com/articles/gex-dex-vex-chex-8-year-backtest-spy-vix-control](https://flashalpha.com/articles/gex-dex-vex-chex-8-year-backtest-spy-vix-control)
