Strategy graveyard / VIX regime switching
Verdict: doesn't work Backtest only
A published VIX regime strategy showed a Sharpe ratio of 1.09. Without its lookahead bug, it matched a fixed mix.
Backtested on SPY, long-term Treasury bonds (TLT), gold (GLD) and T-bills, daily from 2006 to October 2026. Updated 11 Oct 2026.
Short answer: the VIX timing added nothing once a data bug was fixed. I rebuilt a published strategy that moves between stocks, bonds, gold and cash depending on the VIX. As published, it showed a Sharpe ratio of 1.09 from 2006 to 2026, nearly twice that of holding SPY. But the backtest was reading each day's VIX closing price at 10:00 in the morning, six hours before the close existed.
With that fixed, the Sharpe ratio fell to 0.71. Simply holding the strategy's own average mix of the same four assets, never switching, scored 0.70. The switching did one useful thing, avoiding most of 2008, and nothing else.
1.09Sharpe ratio as published, with the lookahead bug
0.71Sharpe ratio with the bug fixed
0.70a fixed mix of the same assets, never switching
+12.3% vs +15.1%a year since 2012: fixed strategy vs plain SPY
What is VIX regime switching?
The VIX measures how much volatility the options market expects in the S&P 500 over the next month. It's often called the fear gauge. Regime-switching strategies use it to decide how much to hold in stocks: more when fear is low or fading, less when it's spiking, with bonds, gold or cash taking up the rest.
The version I tested was published in the strategy library of a popular algorithmic-trading platform, with a machine-learning model on top to fine-tune the stock weight.
The exact rules I tested
Checked every day at 10:00 New York time. The first rule that matches sets the weights.
- Panic and oversold: VIX above the 80th percentile of its last 100 days and SPY down more than 3% over four days → 85% SPY (100% if the model is bullish), the rest gold.
- Calm and stretched: VIX below 13 and SPY more than 5% above its 50-day average → 40% SPY, 30% bonds, 20% gold, 10% T-bills.
- Fear easing: VIX between 20 and its 20-day average → 70% SPY (85% if bullish), 10% bonds, the rest gold.
- Fear rising: VIX more than 20% above its 20-day average → 30% SPY, 40% bonds, 20% gold, 10% T-bills.
- Uptrend: SPY above its 200-day average → 60% SPY (70% if bullish), 25% bonds, the rest gold.
- Otherwise: 30% SPY, 40% bonds, 20% gold, 10% T-bills.
- The model: a random forest retrained at the start of each month on about three years of daily data, using 11 VIX and SPY features to predict whether SPY will be 2% higher in a month. "Bullish" means a probability above 60%.
- What I changed for the honest version: only the VIX data. At 10:00 the strategy may use the VIX up to yesterday's close, not today's. Everything else, including the code's quirks, was kept as published.
- Costs: 0.01% of the value traded each time the weights change.
Results
- Where the bug helped: on days the published version classed as "fear rising", SPY went on to fall 0.58% on average over the next day. Classified honestly, with the VIX known at the time, those days were followed by a gain of 0.08%. The bug was sorting days using information from later the same day.
- The model: with the bug fixed, the rules alone did slightly better than the rules plus the random forest. It added nothing.
- Since 2012: the fixed strategy made 12.3% a year, against 15.1% for SPY and 11.2% for its own fixed mix. Sharpe ratios 0.92, 0.88 and 0.90: effectively the same.
- 2008: the one year switching clearly paid. The fixed strategy made 7%, against −19% for its own fixed mix and −37% for SPY. That single year is most of the difference in worst drop.
- 2022: stocks and bonds fell together, and the fixed strategy lost 16%, about the same as SPY.
- I couldn't reproduce the published headline return even with the bug in place.
Why it failed
- The edge was tomorrow's information. Knowing today's VIX close at 10:00 tells you whether the afternoon will be a sell-off. Any strategy that can see it will look brilliant. It's a common data-alignment mistake when daily data is time-stamped at midnight.
- Diversification isn't timing. Holding stocks, bonds and gold together cuts risk, with or without switching. The fixed version shows how much of the result came from the mix rather than the VIX.
- One crisis carried it. Moving out of stocks in 2008 helped. In every other period the switching made little difference, and in 2022 bonds failed as protection.
- Machine learning can't fix a missing signal. With honest data there was nothing for the model to sharpen.
Did I test it wrong?
Quite possibly. I tested the version of VIX regime switching described above, with settings I thought were fair, and it didn't work. That doesn't prove every version fails. I may have picked the wrong settings, the wrong market or timeframe, or missed a filter that experienced traders take for granted. Many strategies also depend on judgement that a set of fixed rules can't capture.
If you trade VIX regime switching and it works for you, I'd genuinely like to hear how. Tell me the chart and timeframe, what counts as a setup, any filters, where you get in, where your stop and target go, and when you get out. I'll test your version on the same data, the same way, and publish the result here, whichever way it comes out.
Send me your version
What would change my mind
A VIX regime rule that beats a fixed mix of the same assets on risk-adjusted returns, using only data available at the time of each trade, across more than one crisis. If you have one, I'd like to test it.
Questions
What is a VIX regime strategy?
A strategy that changes how much it holds in stocks, bonds and gold depending on whether the VIX (the market's fear gauge) is low, high, rising or falling.
What is lookahead bias?
Using information in a backtest that wasn't available at the time of the trade. Here the backtest read the VIX closing price at 10:00 in the morning, six hours before the close happened.
Does VIX timing improve on buy and hold?
It cut the worst loss from 55% to about 26%, mostly by avoiding 2008. But a fixed mix of the same assets did about as well on risk-adjusted returns, and since 2012 the strategy earned less than simply holding SPY.
Does adding machine learning help?
No. The random-forest overlay made no difference once the lookahead bug was removed.
How it was tested
- Data: daily prices for SPY, TLT and GLD with dividends, the VIX, and the 13-week T-bill rate, 2006 to October 2026. From 2017, SPY was priced at its actual 10:00 level from one-minute data; before that, at the daily open.
- Replication: the published code was followed line by line, including its quirks, and only the VIX timing was changed for the honest version.
- Backtest, not live trading: these results come from historical prices with trading costs included. Backtests usually flatter a strategy, so a failed backtest is a strong "no".
This page records my own testing. It isn't financial advice.
