Back From the Dead to Look Under the Hood of Our Strategy

Hi folks, It’s been a while since I last wrote anything. I had to leave the boat under Zackary’s captainship, not because of a lack of interest, but because I became overloaded in a way I had never anticipated.

My new role as CTO AI at Robotiq came at a moment when Physical AI is exploding, adding another layer on top of everything else. At first, it was manageable. I was still able to post several times in January and into early March. But after a few back-to-back business trips, things started to slip. From that point on, I found myself almost constantly on duty, to the point where I barely had the chance to see my kids, let alone write posts for WU.

The situation has improved a bit recently, as my last semester as a professor has now come to an end. Although I announced in January that I was leaving the university to return to industry, the reality is that by the time I made that decision, the student enrolment period was already over and my two winter courses were already full. So the university asked if I could still teach that semester. I also had two PhD students close to graduation, which added even more to the workload. To give you a sense of how packed my schedule had become, on each day I was teaching, even the 20-minute break in the middle of a three-hour class was booked with meetings.

That situation was unusual, and it should continue improving into the summer. I expect to have more time to write frequent updates again, and probably to work on some of the things I will discuss later in this post.

But for now, I want to discuss the hedge signal we had on March 30. The fact that it triggered near the bottom is not ideal, but it does not bother me that much, particularly since the strategy re-entered rapidly. What I found more concerning was this:

These numbers are for QQQ not SPY. Still, seeing the strategy print a dot outside the range of previous trades made me ask myself: Is the hedge strategy broken?

This post is a summary of what I found after taking a deeper look under the hood of the strategy, its recent behavior, and the stats behind it.

What the hedge is, and what it is not

First, let’s recap what the hedge strategy is designed to do.

When we built it, the goal was to avoid being caught in the biggest corrections. Those are the ones that are the most painful at the index level and can take a long time to recover from.

Of course, the ideal scenario would have been to build an algorithm that could see every meaningful dip coming. But when I started working on this in 2021, without AI, I quickly came to the conclusion that this was simply impossible. And even today, five years later, with a much more AI-empowered version of myself, I still believe it is impossible to cancel market volatility. Volatility is the price we pay for returns.

On the other hand, it quickly became clear to me that stepping out of something that is morphing into a major decline is relatively easier. I am talking about the GFC, the fall 2018 correction, the Covid drop, the 2022 bear market, and the 2025 tariff correction. These large drops tend to leave a very specific imprint, with multiple outlier signals showing up around the same time. Just last year, in February 2025, three different components of our hedge signal fired strongly within a few days.

So hedging every small drop seems impossible to me. Building a strategy that reacts to massive drops is much easier. The tricky part lies in between: the not-so-small corrections, where some are on the side we should not hedge, while others are on the other side of the fence.

I have often said that the first version of the hedge strategy was used very successfully during the October 2021 correction. But when I started working with the IOFund in January 2022 on a dedicated version for them, that same October 2021 correction became the prototype of the ultimate correction we should not hedge.

We had two reasons for choosing it as the frontier.

First, we both agreed that it was painful, but not the type of correction that creates deep and lasting damage to a portfolio. Second, I had already run a lot of simulations by then, and any strategy that led us to hedge that correction, even successfully, produced a lower return and hit rate than a strategy that ignored it.

To give you some perspective, here are a few stats on SPY about that October 2021 correction. This will be useful later in the post.

·       Biggest intraday move: -2.83%

·       Biggest gap: -1.49%

·       Biggest close-to-close move: -2%

·       Overall drop from the highest daily close to the lowest daily close: 5.48%

·       Overall drop including intraday price action: 6%

This does not mean the strategy will never hedge before a specific drawdown threshold, or when volatility is lower. But a correction with a similar signature, or a less brutal one, should generally be ignored. A correction with a stronger signature, on the other hand, should probably not be ignored on average.

 A Correction Near the Frontier

But what we had this time was exactly in the middle. It was not a weak and quick move, but it was not a major correction either. It was a months-long correction with a signature very similar to our frontier example.

·       Biggest intraday move: -2.15%

·       Biggest gap: -1.71%

·       Biggest close-to-close move: -1.71%

·       Overall drop from the highest daily close to the lowest daily close: 8,8%

·       Overall drop including intraday price action: 9%

Since a graph is worth a thousand words, here is how these stats compare with our frontier correction from October 2021, and also with the much larger correction we had last year around Liberation Day and the tariff-related drama.

So, as I said above, this was exactly the kind of correction that sits in the difficult middle zone. It was not a violent high-velocity drop, but it was long and painful enough to raise the question of whether it should be hedged or not.

In these cases, the strategy is basically forced to answer a very hard question: is this the end of the correction, or is it about to morph into a much more brutal C wave?

Sometimes, in these frontier cases, the strategy gets it right. October 2014 is a good example, where the signal triggered around the midpoint and avoided the last brutal wave.

But in other cases, like April 2018, the market did not go much lower after the exit point.

So the setup itself was not absurd. It was exactly the kind of borderline correction where the decision is hard.

But despite that, the strategy still printed an outlier...

An outlier?

The hedge strategy has a hit rate of around 88% over the last 20 years. Since we only go market neutral, and not short, a successful trade simply means that the strategy sold SPY on a hedge signal at a higher price than where it bought it. This is true regardless of the price at which we re-enter, which is a separate question from how much the strategy beats buy-and-hold on the same asset.

That 88% hit rate is the overall number, but it varies depending on the period. If we look at the period I used to design the strategy, up to 2019, the hit rate was around 78%. Since the strategy went live, 92% of the trades, or 12 out of 13, were good trades.

So even though the last trade was negative, it was only the first live trade where the strategy exited at a lower price than where it entered. That is still well within what the historical stats suggested could happen. And when we eventually get a more choppy market, like in 2013 or 2015–2016, I expect the strategy to produce more of these losing trades.

So in that sense, the strategy is not broken.

But as I said, what worried me was not simply that the trade was a losing one. It was how far the exit point landed from the entry point compared with previous trades. I would not have been too concerned by a losing trade that still sat inside the point cloud of previous losing trades, these have to be expected.

But seeing it well below that cloud is a different story.

The end result of a trade depends on a few simple things: the starting point, how much SPY rose during the trade, and how much it retraced before the hedge signal triggered. The combination of these factors is what ultimately makes a trade a good one or a bad one.

For example, one of the best trades in the strategy’s history ran from March 16, 2009 to August 2, 2011. During that period, the strategy locked in a 65.42% gain, which is enormous. (Imagine having ridden that move with UPRO or TQQQ...)

That being said, the trade actually peaked at 80.83% before retracing a massive 15.41% into the exit. That was one of the worst retracements from the high that the hedge strategy has ever experienced.

Retracement is normal. Although the strategy has, on a few anecdotal occasions, sold near the top, it is not designed to do that. Its function is to exit when strong signs of danger appear. The average retracement before exit is around 4%, so seeing the 2011 trade give back 15.41% on SPY was a strong outlier.

By pure coincidence, the strategy delivered almost the exact same gain on the very next trade, at 65.19%. But this time, the trade ended much better, as it only saw a 4.42% retracement from the high before the hedge signal. So retracement is a very important part of the equation.

Retracement was also the main factor in the losing trade from September 3, 2015 to January 7, 2016, where the strategy exited SPY 0.77% lower than where it had entered. That trade did have a non-negligible run in between, but the strategy retraced 9.67% before exiting.

In the case we saw in March, the retracement was higher than average at 8.8%, but still lower than several other trades in the strategy’s history. The main culprit behind the outlier losing trade was that we simply had almost no run-up during the four months that the trade lasted. This is the unusual part.

As we said several times in early 2026, after the huge run-up the market had in 2025, it entered a very long consolidation starting in October 2025. As a result, after the strategy successfully re-entered the market following the November correction, SPY at is max only closed 1.75% higher than our November 28, 2025 re-entry price over the four months that followed. That is something we have not seen often.

So the outlier was not caused by an extreme retracement alone. The retracement was larger than average, but not unprecedented. What made the trade stand out was the combination of almost no prior run-up, followed by a larger-than-average pullback. Together, those two factors pushed the trade well below the historical cloud of previous losing trades.

Not an outlier, but still an outlier

Despite that non-negligible losing trade, if we were to close the current trade today, the strategy would still have outperformed buy-and-hold by about 22% on SPY over the full period it is live.

That said, it is underperforming by roughly the same amount since the start of the new bull market (end of 2022).

This does not surprise me. In the very first analysis I did on the strategy, I pointed out that the results can vary a lot depending on the three-year window we look at.

In bear markets, the strategy is almost always positive and tends to greatly outperform buy-and-hold. In strong markets with sharp corrections, like 2018–2020, it can also outperform buy-and-hold by a wide margin. But in massive bull markets, or in choppier environments, the strategy often slightly underperforms. This has happened in the past with the strategy, and it is exactly what we are seeing here.

After all, when the trend is very bullish, drawdowns are shallow, rebounds are strong, and there is simply not much to hedge against. These stats could improve quickly during the next major correction. If the market only keeps trending up, they may not, although I doubt that scenario given that the PE ratio is now above 32...

But if I had to point to one specific factor that made this period even more difficult, it would be the highest market concentration we have seen in roughly 50 years, driven by the Magnificent Seven. We have to go back to the early 1970s and the Nifty Fifty era to find a similar phenomenon. For those not familiar with the term, I am referring to the record amount of capital that flowed into the top seven stocks: Apple, Microsoft, Nvidia, Amazon, Alphabet, Meta Platforms, and Tesla.

After the post-Covid speculative bubble deflated, and after the fastest Fed rate-hiking cycle in decades pushed investors toward cash-rich companies, the AI narrative then added another layer to this concentration. Those seven companies attracted a record share of capital, reaching at one point about 34% of the total S&P 500 valuation.

Why does that matter? Because the strategy is built around the index, and the index is made up of 500 stocks. The S&P 500 has always been pushed by its largest companies, but rarely to that extent. At some point, SPY’s price action became heavily tied to the trajectory of these seven stocks. This helped us last summer lock in a near-record profitable trade, as these forces drove nearly alone SPY and QQQ to a considerable new high. But at the same time, it also negatively affected the strategy.

One of the cornerstones of the strategy is the market-breadth dataset. I strongly believe that someone could almost hedge using this dataset alone and still avoid every major drop. But extreme market concentration distorted this lens.

In 2024 and 2025, there were a few moments where breadth deteriorated heavily while the index itself showed very little damage. Last November was a good example. The pullback was very small at the index level, and under normal conditions, the strategy would likely have ignored it. But under the hood, starting in early October, we were already seeing massive damage in riskier and more speculative stocks.

Investors exposed to those stocks were probably already down around 30% when the hedge signal triggered, even though SPY was only about 3.5% below its all-time high.

At the beginning of this year, I expected to start seeing some return of the pendulum, and this is exactly what happened. While the index went nowhere and the Mag Seven started to retrace, market breadth improved.

This continued into early March. The index started to move meaningfully lower, but market breadth mostly went sideways. It looked like the Mag Seven trade was unwinding, which made the hedge trigger late.

Since the market bounced back, we are now in a more balanced situation, where both large caps and smaller stocks have recovered in healthier synchronism (Just to be clear, the synchronism is healthy; the speed of the bounce is another story).

Because of that, I expect the market-breadth dataset to start behaving more in line with its longer-term history, rather than the distorted behavior we saw over the last two years.

I also asked Claude Cowork, using Opus 4.7, to analyze that dataset and see whether it could find a better threshold than the one I found years ago. After burning through a lot of tokens, it came back with what it presented as the optimal setup. I then gave it the thresholds we currently use, and it realized that our strategy was actually better than the one it had found.

So, 1–0 for the human against AI. But in reality, Claude arrived at something very similar (and anecdote, it produce itself a "tunemap" for optimizing!) to what we already had, which reinforces my view that we probably should not change this part of the strategy. Here is the CSV of the data in case you want to give it a try yourself.

Maybe this is not the part to change, even if it was the culprit behind the late hedge signal this time. But there are still things I would like to improve, and the next, and last, section is about that.

Is the hedge strategy broken?

All the stats we have today point to the same answer: definitely not. This is true even though we have been in an incredible bull market since 2023, one that has favored only a handful of players in a way we had not seen since the Nifty Fifty in the early 1970s. As we said in the previous section, this kind of market concentration does affect our market breadth indicators. It made them a bit early in some cases, like in November 2025, where I would have preferred not to see a hedge signal, and late in others, like in March of this year.

But this behavior is still within the kind of variability we should expect from a strategy that was designed on the 2005–2019 period and has been running live since the end of 2021. More importantly, nothing I have seen has changed my belief that the strategy would still raise a flag before the worst part of a major correction, which is the most important part. A more balanced market should also make those indicators behave much more in line with their long-term history again.

That being said, if you remember my usual New Year’s Eve post from last January, improving the hedge strategy was one of the things I said we wanted to work on in 2026. What happened in March is exactly the kind of case that illustrates why I thought this was important.

The first reason is simple: we now have a longer live history that we can reuse in the design phase.

When I worked on the strategy in 2021–2022, we did not use all the available history. The reason was to avoid overfitting. You want to keep an out-of-sample period to test whether the strategy really works, or whether it was simply fitted too tightly to the past.

At the same time, you also want to use recent history, because markets evolve slowly but surely. For example, the market’s aversion to speculative, non-profitable assets since 2023 is very different from what we saw in 2019–2022, when a simple narrative about how a company would change the future was often enough to drive the stock price higher.

So the healthy choice is a mix: use recent history, but still keep enough out-of-sample data to avoid fooling ourselves. Now, five years after I started working on the hedge algorithm, the healthy choice is definitely to re-tune the strategy using the additional four to five years of data we now have.

But there is another reason why I think it is time to reopen the hedge strategy, and March illustrated it very well.

From the outside, we often see a karateka reaching black belt as the point of mastery: that person knows karate.

But in reality, if they continue training for years, they keep improving way past that point.

It is the same thing in science. When I came back from my postdoc at Stanford to start my career as a university professor, even though I had started teaching, I certainly did not stop learning. I am way more knowledgeable in robotics today than I was then.

I can say the same thing about the work I have done trying to build algorithms for finance. These last five years gave me a lot of experience. I probably wrote thousands of pieces of code and tested thousands of ideas beyond what you can see on the Wealth Umbrella dashboard. During that time, I also tested thousands of datasets, particularly while working on WU Advanced.

As I said when we launched it, by 2025 I felt I had exhausted most of the data options available on the market. And I still consider our downtrend exhaustion indicators, what I used to call the “buy the dip” indicators, to be the best work we have ever done.

Did you see how well they have worked since going live? Particularly DE1, which has nailed the B bounce (in large corrections) or the bottom, almost every time.

And again this year.

At the moment, these very powerful datasets, one of which was actually found by AI, are not part of the hedge signal at all. That means that whatever these downtrend exhaustion indicators are trying to tell us about where the market is going, the hedge strategy is completely blind to it.

This is exactly what happened in March. Our DE2 signal had just sent a strong orange signal two days before the hedge kicked in, which was a clear contradiction.

Well, maybe not a total contradiction. The hedge strategy is not opportunistic. It does not try to identify bottoms. It simply raises a flag when it starts to see risk. DE1, DE2, and DE3 are different. They look for reversal patterns. And rising risk often creates the conditions for a reversal. That is what DE2 and DE3 were seeing around the hedge signal, and DE1 confirmed it the day after the hedge.

This is why, after pledging last year that we would no longer ignore a full hedge signal, we only sold the leveraged part this time. If it had been only us, we probably would not even have sold that part. But we wanted to stay consistent with what we had said, and at least deleverage. But.... The difference this time was that we had these new indicators, and we trusted what they were showing. What we did not have yet was an integrated and backtested framework that combined them with the hedge strategy. So we made a compromise.

But from an investor’s point of view, having a system that can take all of that into account would make much more sense.

I would also like to include the color signal from our Margin Risk indicator, which has proven to be very accurate in the past. I know some of you have already built your own systems around the combination of these indicators, and I personally combine them manually as well. But I think having a backtested and integrated approach would be very useful.

Finally, I have one last motivation for updating our strategy, and it is related to AI.

Over the last year, AI models have made a tremendous jump in how they handle code and system design. They went from being helpful assistants that could write chunks of code to speed up development, to near-autonomous systems that can work for hours on complex projects.

I think this will be incredibly powerful when trying to optimize a more complex hedge strategy. For those of you who have used TuneMap, you know that it is easy to visualize the maps when there are only two or three parameters to tune. But the upgraded hedge strategy I have in mind could use around 16 datasets, each with multiple thresholds. For a human, that becomes very challenging to tune properly. For an AI-augmented human, it becomes a much more manageable task, and one that could potentially lead to very strong results. That is something I did not have access to in 2021–2022, when most of the strategy was coded.

Conclusion

So, even though the last trade was unusual from some angles, it was still well within what we should expect from the strategy when we look at it from most perspectives on a longer term.

But, as we already said in January 2026, we think it is now time to spend more energy on improving it. This is what Zackary will spend time on this summer. And I need to say it again: Zackary is one of the best engineers I have ever met in my career. Around that time, I should also start to have more availability, and I expect to help him with this. You should also hear more from me again.

The month ahead will still be incredibly loaded, and this will probably continue to keep me mostly away from WU for a bit longer. This is not exactly what I had envisioned when I jumped back into industry, but it will not stay like this forever.

That being said, on the positive side, even if I have not been writing updates as frequently, working at the current heart of AI and interacting with many of the major companies in the field is teaching me an incredible amount. I am convinced that this will eventually become very beneficial for WU as well. It is too soon to explain that part in detail, but later this summer, I promise I will tell you more about it (this post is anyway already long).

In the meantime, Zackary should soon roll out a completely new version of our platform. It will fix bug and improve the visuals, but more importantly, it will do a much better job guiding people on how to read our different indicators, how they are related, and how they should be used.

Stay tuned.