Help us double down on what's working, instead of guessing. Takes 5 seconds, totally optional.
VIX Term Structure Inversions Since 2018: What Backwardation Actually Signals
Every VIX term structure inversion from 2018 to 2026, catalogued from 413 weeks of point-in-time data. The market spent 14.5% of weeks in backwardation, median VIX during inversions was 23.8, and the surprising result: inversion carries almost no directional edge (median forward returns nearly identical) but forward return dispersion jumps 60%. Backwardation is a position-sizing signal, not a direction signal.
The VIX term structure spends most of its life upward-sloping: spot VIX below 3-month and 6-month VIX, the market charging carry for future uncertainty. Inversion - near-term vol priced above longer-term - is the curve's alarm state. Everyone quotes it; almost nobody has measured what it actually predicted, because doing so needs the full history of the curve as it stood at each moment. The historical summary endpoint carries the VIX complex (VIX9D, VIX, VIX3M, VIX6M, slope, and a contango/backwardation label) for every replay date, so we sampled all 413 Friday closes since May 2018.
Method
Sample: 413 Friday closes (Thursday on holidays), 2018-05-04 to 2026-03-27
Signal: the API's vix_term_structure.structure label - backwardation when the near end of the curve (VIX9D vs VIX) inverts
Outcome: SPY return over the following 4 weeks, plus the dispersion of those returns by regime
Episodes: consecutive inverted weeks grouped into episodes with start/end dates
The Base Rates
Weeks in backwardation
14.5% (60 of 413)
Distinct inversion episodes
37
Median VIX while inverted
23.8
Typical episode length
1-2 weekly samples; the exceptions are the famous ones (Covid: 6 weeks)
The Test: Direction vs Dispersion
4-week forward SPY return
Inverted weeks
Contango weeks
Median
+1.45%
+1.69%
5th percentile
-10.55%
-6.56%
95th percentile
+11.46%
+6.33%
Standard deviation
7.01%
4.39%
Read the medians first: the naive rule - "sell everything when the curve inverts" - would have accomplished nothing on average, because inversions cluster mid-panic, and mid-panic is where violent rebounds live too (both distribution tails widen, in almost perfect symmetry). What inversion does reliably price is range: the next month's outcome distribution is ~60% wider in backwardation. The curve is not forecasting direction. It is forecasting the size of whatever happens next.
The practical translation: backwardation is a position-sizing input, not a direction call. Halving position size when the curve inverts keeps the same expected return profile at roughly the contango-era risk. That is a materially better use of the signal than flattening exposure and missing the rebound half of the distribution.
The Episode Catalog
All 37 episodes are in the dataset; the ones every trader will recognize:
Episode (weekly samples)
Duration
What it was
2018-10-05 to 2018-11-09
5 weeks
October 2018 selloff
2018-12-21 to 2019-01-04
2 weeks
December 2018 capitulation
2020-02-28 to 2020-04-03
6 weeks
Covid crash - the longest inversion of the era
2022-01-21 to 2022-12-16
6 separate episodes
The bear-market year: the curve inverted and re-inverted all year without ever staying inverted
2024-07-26 to 2024-08-02
1 week
Yen-carry unwind / August 5 vol spike
2025-03-07 to 2025-04-17
2 episodes, 5 weeks total
Tariff shock
2026-03-06 to 2026-03-27
4 weeks, ongoing at sample end
The current stress regime
Two structural observations from the full catalog. Most inversions are short - one or two weekly samples - and resolve back to contango without a crash; the curve throws more alarms than there are fires. And 2022 shows the signal's other failure mode: in a grinding bear, the curve flickers in and out of backwardation for a year, whipsawing any binary de-risk rule. Duration, not occurrence, is what separated the historic events (Covid: six consecutive weeks) from the noise.
Honest limitations. Weekly Friday sampling merges intraweek flickers and can split one economic event into two episodes (the 2025 tariff shock reads as two). The structure label keys off the near end of the curve (VIX9D vs VIX), which inverts earlier and more often than VIX vs VIX3M - a stricter definition would show fewer, longer episodes. Forward-return stats are descriptive; inverted weeks cluster into a handful of events. The macro series is EOD, applied at each replay date.
Rebuild This Series
# VIX complex (VIX9D/VIX/VIX3M/VIX6M, slope, structure label) at any date
curl -H "X-Api-Key: YOUR_API_KEY" \
"https://historical.flashalpha.com/v1/stock/SPY/summary?at=2020-03-20"
Backwardation means near-term implied volatility is priced above longer-term implied volatility - the VIX term structure slopes downward instead of its normal upward (contango) shape. It signals the market is paying a premium for immediate protection, which historically happens during active stress. From 2018 to 2026 the VIX curve spent about 14.5% of weeks in backwardation, with a median VIX of 23.8 during those weeks.
The 2018-2026 data does not support outright selling: median 4-week returns after inversion (+1.45%) were nearly identical to contango weeks (+1.69%), because inversions cluster mid-panic where sharp rebounds also live. What changed dramatically was dispersion - forward returns were about 60% more volatile. The historical evidence favors reducing position size during backwardation rather than exiting, keeping exposure to the rebound half of the distribution.
37 distinct episodes in eight years - roughly four to five per year - but most last only a week or two and resolve without a major event. Long inversions are the rare, information-rich ones: the Covid crash produced six consecutive inverted weeks, the longest of the 2018-2026 era. Duration separates real regime breaks from routine stress flickers.
Conclusion
The VIX curve's alarm is real but widely misread. Inversion does not say "down" - eight years of episodes say it never reliably did. It says "bigger": both tails of the next month's distribution widen by more than half, which is exactly the information a position-sizing rule needs and a direction rule wastes. Treat backwardation as a volatility forecast for your own P&L, watch episode duration rather than occurrence, and let the historical curve data keep you honest about what it has and has not predicted.
