The Bear Mode Rework

As part of the rework of our hedge signal that began this summer, our first step was to take a closer look at the bear-market component. Vincent already touched on this in the project explanation blog post, but for the sake of completeness, here is a short reminder of how it works.

Our hedge signal operates in two distinct modes. In their original work, Jennifer and Vincent found that the “standard” metrics used by the strategy did not perform reliably during bear markets. Bear markets are largely driven by fear, and fear-driven markets tend to follow momentum more than anything else. For this reason, they decided to use two distinct strategies depending on the type of market environment.

It was therefore important to begin our work by making sure that our method for detecting the start and end of bear markets was as robust as possible. Once that was established, we conducted a full audit of the bear-market strategy itself, questioning every component of the methodology as well as each threshold selection.

The bear-market section of the algorithm goes beyond the Bear Signal you have access to on TradingView. It also includes the logic that determines how the strategy triggers within a bear market to capture momentum. Those of you who were with the IO Fund in 2022 may remember that this part of the strategy relied heavily on phase angle and the VIX. So let's start with an overview on how the Bear Market strategy works before exploring what we changed and the results.

The Bear Market Strategy

The Bear Market strategy has two distinct mechanics. The first determines when a bear market starts and ends, which controls when the strategy switches into and out of Bear Market mode. The second determines how we enter and exit positions while Bear Market mode is active.

Let’s first look at how we define the start and end of a bear market. The obvious starting point is the standard -20% drawdown. Vincent challenged this threshold in the past, and I challenged it again, only to reach the same conclusion: statistically, it works surprisingly well. Set it slightly lower and you start classifying some non-bear episodes as bear markets. Set it slightly higher and you miss some actual bear markets. Why 20% works so well is unclear. Maybe it has become a self-fulfilling threshold: as soon as the market reaches -20%, news networks are suddenly filled with banners and panels announcing that we are officially in a bear market. Perhaps that fuels the fear, perhaps not. Either way, the data supports it, so we stick with -20% as our definition of a bear market start.

Defining the end of a bear market is more complicated than simply climbing back above the -20% line. In long or brutal bear markets, it can take years to recover to that level. In others, a strong rally can briefly cross it before the market turns back down and makes another low.

In our Bear mode, there are three ways the strategy can call the end, and they all share the same philosophy: each looks for something that no bear market rally in history has ever produced. The first two look at how far the market has recovered. We measure this in two ways: the rise from the latest market bottom, shown by the blue line in our Bear Signal indicator, or how close the market has recovered toward its pre-bear peak. The rise from the bottom is especially useful during deep bear markets, when recovering the previous peak could still be years away. The second measure works better for shallower bear markets, where waiting for a large rise from the bottom could keep us in Bear mode longer than necessary. Both have remained statistically sound across all historical bear markets; the only difference is which one calls the end sooner.

The third way asks not how far the market has recovered, but for how long the recovery has held. The measure behind it is what we call the BWR, a ribbon of EMAs of different lengths. During a bear market the ribbon is inverted, and the sharp rallies bears love to produce are simply too brief for the longer EMAs to fully turn back around — the ribbon never re-opens before the rally dies. Only a recovery that lasts can pull the entire ribbon back into bullish alignment and stretch it to a spread that no rally inside a bear market has ever produced. That gives the strategy a way to call the end of a bear when the recovery is real but not yet steep enough to have reached the rebound or reclaim thresholds, moving the re-entry up by a few days to a few weeks in the bears where this call wins the race.

Finally, being in a bear market doesn’t mean there are no tradable rallies. It simply means that caution must come first, as fear-driven markets can move very quickly in either direction. This is why the strategy doesn’t simply sit out the entire bear market. Instead, it looks for conditions where strong rallies are likely, while remaining ready to exit quickly under strict safety rules. Entries must come from clean, strong upward momentum. This is what our backtests showed worked best, and we saw it work in real time during the 2022 bear market.

Two conditions detect this, the first one is our Phase Angle indicator which by default is tuned to catch bear market rallies. One thing to note is that there is a veto to this momentum entry to prevent entering during high or rising volatility. A strong momentum push while the VIX is too high will not trigger an entry as those are usually coin flips and during bear market we prefer to lean towards safety. As we can see in the 2022-2023 bear market, the Phase Angle produced good entries and exits, but the most dangerous ones did not trigger an hedge signal entry due to the veto.

The second one is monitoring the unusually strong push with low volatility inside a bear market that our Bear Signal indicator is designed to detect. Vincent analyzed the rallies within all historical bear markets and found that sustained rallies tend to share a specific signature, as shown in the image below. That signature is a strong and consistent upward push characterized by relatively low intrinsic volatility. It sends a clear message that market participants are rushing back in. 2003, 2009, 2018, and 2020 are good examples. This type of push has appeared in every bear market in our historical data and was almost always followed by a strong rally when it didn’t simply mark the beginning of a new bull market.

All of those conditions were accompanied by a handful of vetoes in addition to the VIX one already mentionned. After investigation, those veto could not justify their presence and thus have been removed reducing the degrees of freedom of the strategy without negatively impacting its historical performance.

While the risk in a bull market is being out of the market, in a bear market it is the opposite: you don’t want to get caught riding the elevator down. Accordingly, our strategy is deliberately asymmetric. It requires a strong upward push to enter the market, but only a mild reversal to exit. The exit condition uses the Phase Angle with a deliberately tight threshold, allowing the strategy to get out at the first sign of downward momentum. This prevents the strategy from getting caught in another major leg of the drawdown.

What the rework did

Now that we've went over the specifics of our bear market mode, let's dive into what this rework ended up doing. One of the reassuring conclusions from this work was that Vincent had done a pretty good job five years ago of defining strong, statistically sound thresholds. Most of them held up well and required no adjustment. The main changes therefore came from simplifying the strategy rather than fundamentally altering it. We removed components that, with the benefit of additional data and testing, proved less useful than they initially appeared. We also made a few small adjustments to threshold selection to improve consistency across the strategy. Here's a short list of the results:

What did not change:

• The -20% bear start condition — the strategy steps aside on exactly the same days as before.

• The riseup bear-end call, the bear bounce push, the phase angle entries and exits, and the VIX conditions around them — all kept their existing thresholds.

What changed:

• Two bear end paths were tuned for better performance: the pre-bear peak recovery and the final push signature.

• A patch was added for a failed bear end call: In a specific condition where a bear end would be called without being the real bear end, the strategy could remain stuck in Bull mode. This is very unlikely as it requires a specific setup to occur, but to be safe we have added a fallback option to re-trigger a bear start upon declining below the previous bear bottom.

• A few thresholds were nudged to more robust positions inside their plateaus, with zero change to the historical record.

What was removed:

• A handful of minor entry vetoes and one redundant bounce-entry leg that could not justify themselves over the full backtest.

• One condition that proved practically impossible to trigger in any realistic scenario.

In short, the work can be summarized more as a cleanup than a complete rework. We have made things simpler, more robust and coherent while slightly improving historical performance. To measure this improvement, we have tested the bear mode strategy in isolation. This means that during bull markets, the strategy remains fully invested and once a bear start is called, already minus 20% down, the strategy exits and follows the bear mode calls. Starting mid 2003, right around the time all of the required datasets become available, the annualized return of the strategy increases from 12.4% to 12.6%, compared with 9.4% for simply holding the market through everything. This is a slight yearly improvements, but on a set of conditions that are now simpler and better tuned.

The next part of this blog post will go in more details about the work done with more data, examples and figures. If this is not something you're interested in, you can definitely skip this as you already have the big picture of what was done. However, if that's something you like, let's jump in!

What this is all about

This work started with a pretty simple question: if we put every rule in our bear strategy on trial today, would they all survive? Vincent built this part of the algorithm almost five years ago and spent several weeks testing it extensively. The work was long and tedious, mostly because we wanted to make sure there were no mistakes and that every conclusion was supported by the data. What has changed since then is not so much the data although we do have longer history now, but the tools available to us. With AI coding agents, work that previously required months of writing one-off scripts can now be done in a few weeks. For the first time since the strategy was launched, we could therefore afford to question pretty much everything rather than simply revisit individual pieces when needed. So that is what we did. We rebuilt every rule in a research environment, replayed it against history, and essentially asked two questions: Does this rule really deserve to be there? And if it does, is the threshold we selected still the right one?

Before starting, however, we established two important ground rules. The first was that every rule had to prove that it was useful on its own before we looked at how it interacted with the others. Why does this matter? Imagine the worst possible way to build a strategy. You have thirteen historical bear markets, so you create thirteen different rules, with each one specifically tuned to catch one of them. Put everything together and your backtest could look absolutely fantastic. The problem is that you have learned almost nothing about the common denominators of bear markets, and your strategy will likely fail when bear market number fourteen comes along.

This is what we call overfitting in machine learning. The more degrees of freedom—or adjustable parameters—an algorithm has, the easier it becomes to overfit. Each additional parameter gives you another opportunity to tune the algorithm to capture a specific anomaly in the historical data rather than something that will reliably repeat in the future. The thirteen-rule example above is an extreme illustration of exactly that problem. Below is a visual analogy of what a strategy does. It tries to split the good trades from the bad ones. If we have a lot of parameters, we can make the line wiggle its way around specific points like the green line shown below. With few parameters (represented by the black line), the line must always keep its shape and can only be moved around making it much harder to find a placement that perfectly fits every single points. The black line will statistically hold much better to future data than the green line.

Obviously, our strategy was never built that way, but the example illustrates the danger. A rule that only works because it patches one or two specific situations missed by other rules can be a smaller version of the same problem. So we tested every rule independently. This does not mean a rule needs a 100% success rate to survive. Very few useful market indicators would pass that test. What we wanted to see was that a rule worked consistently across history rather than being useful only in a handful of very specific cases.

Our second ground rule was to judge each rule based on the mistake it absolutely cannot afford to make. Of course, we care about the obvious question: does this rule ultimately help protect capital or generate better returns? But different parts of the strategy have very different jobs, and therefore very different ways of failing. The rules that take us out of the market cannot repeatedly miss real bear markets. The rules that trade momentum inside a bear should not buy just before another major leg down or sell just before the market takes off. And the rules that tell us the bear market is over should not give us the all-clear while the bear is still very much alive. In other words, we did not grade every rule using the same test. We graded each one according to the risk associated with the job it is supposed to do.

The data

For price-based conditions, we used every trading day of the S&P 500 since March 4, 1957, when the modern 500-stock index was introduced. There is market data available before 1957, but it is reconstructed from the older 90-stock S&P index. It also comes from a somewhat different market, including six-day trading weeks and significantly higher volatility. We therefore decided that 1957 was the cleanest starting point for testing price-based rules. Not every metric goes back that far, however. Some of the datasets used by the strategy only became available in the early 2000s. As a result, individual price-based rules can be tested over nearly 70 years of history, while the complete bear-mode strategy is better evaluated over the more recent period for which all of its inputs are available.

There is one other methodological decision worth explaining. For this bear-mode rework, we deliberately did not keep part of the historical data hidden for final validation. In machine learning, this is normally called a holdout or out-of-sample period, and under most circumstances it is something you absolutely want to have. The problem here is sample size. There have only been thirteen bear markets in the entire dataset. If we split those into training and validation groups, we end up with two samples that are simply too small to tell us much with confidence. We therefore kept all thirteen bear markets available for the analysis. Of course, using all the data creates its own risk: overfitting. We were very conscious of that, which is why we put other safeguards in place to keep ourselves honest.

Beyond sample size, there is a second reason we were comfortable skipping the holdout: degrees of freedom. The bear rules are mainly momentum driven and of a relatively simple construction, a threshold or two sitting on top of a simple measure, so there is not much in them that can memorize the past in the first place. And the place where overfit risk genuinely lives is one level up, where many components get wired together; that layer has its own separate guardrails. So every bear market stayed in one bucket, every event weighed the same, and robustness was read from different questions: does the conclusion hold on both the older and the newer half of the record, and does it survive if you nudge the parameter a step in either direction. Those are questions our data can actually answer.

We also plan to dedicate a full blog post to overfitting and holdout as part of this rework. It is a topic that comes up regularly in questions from our members—and for good reason. It is probably one of the most important concepts to understand when evaluating any strategy built from historical market data. Until we get to that post, however, you may just have to trust us on this one.

How it went

So how does one actually challenge a threshold? The method is worth showing, because we ran it on every threshold in the block. The test itself is simple: slide the threshold across all its possible values and replay history for each one. You would expect the results to change gradually as the value moves. They do not. History reacts only at a handful of exact values, where one old episode flips from caught to missed, and between those values sit wide flat zones where every choice produces the identical past.

We call the flat stretches between them indifference plateaus. Every value inside a plateau produces the identical history, so any placement inside it holds the same statistical robustness. What this picture teaches is that a backtest can only tell us in which plateau the threshold should sit, but it cannot tell us where in the plateau would be better as there is no statistical differences between one end to the other. So how do we choose the spot logically? By asking what a mistake costs on each side. When both sides hurt equally, stand in the middle and keep the maximum margin against both. When one side is expensive and the other is cheap, forget the middle and stand as far from the expensive side as the benefits allow:

Where to stand inside the plateau follows the cost asymmetry shown above, with one refinement that matters in practice: how the measure itself moves. Some measures spike, i.e. they push up toward a peak and fall back. If that peak lands halfway into a wide plateau, a threshold parked at the far edge is never reached and the signal is missed entirely. An example of this is any measure that computes a Z-Score. Even if the underlying data remains high for a few days, its Z-Score will decrease as a high value will no longer be out of the ordinary compared to recent history.

Other measures climb steadily: they will cross a higher threshold anyway, just a few days later, so pushing toward the far edge costs very little. The shape of the measure decides how much of the plateau you can actually spend. A classic example of this would be looking at how much gains the market has made from the bottom of a bear market. If we really are at the end of a bear market, waiting for a few more percents can be justified if it yields a safer entry.

One of the bear-end confirmation rule is the cleanest asymmetric case we have, so here it is with the numbers. It had already cleared the standalone bar; what remained was placement. This specific metric is one that climbs steadily, so being more patient cannot create a failed detection, just a slightly later one, and the other re-entry rules cover the misses. We moved its threshold as far from the expensive side as the benefits allowed, giving up two historical catches that measured out to zero days of value, because the other rules had beaten this measure to the punch both times. In the end, the extra safety cost about five days of added waiting spread across four bear markets, and it bought roughly 60% more distance from the worst false reading the record ever produced. Here is that decision on the real data, with the axis values withheld:

This is not just theory, and last spring gave us a perfect example. On April 24, 2025, one of our re-entry triggers missed its threshold by only a few thousandths. That tiny miss could have ended up being costly: the signal did not finally trigger until May 12, when the market was already 6.8% higher. In fairness, other conditions would have blocked the April signal anyway, but those are conditions that this audit ultimately removed because they could not justify their existence across the full backtest period. We will come back to that in a moment.

What was particularly interesting is that lowering the threshold by just 0.007 would have captured the April signal without changing the historical backtest performance at all. Both values sat on the same performance plateau, meaning there was no statistical reason to prefer the original, more restrictive value. This is exactly the type of unnecessary precision we want to avoid in the future: if two nearby thresholds perform equally well historically, we would rather select the one that gives the strategy a little more robustness instead of pretending that one exact number is somehow special.

The threshold, however, was only half of the story. The other half is one of the conditions that was blocking that April signal anyway, and its removal is a good picture of what the audit's cuts looked like in general.

The condition was a veto on the re-entry side built around short-term volatility. The idea behind it was insurance on the signal itself: elevated short-term volatility points to acute short-term stress, and that stress can resolve in either direction. The situations it flagged were coin flips by design. We added it years ago to lean toward safety on those coin flips, accepting that it would occasionally hold us back from a good entry in exchange for protection when things turned sour.

Over the entire backtest, the veto had never hurt a trade in a major way, but it had never helped one in a major way either: its wins were a handful of small positive nudges. Its true purpose was to reduce the exposure to the volatility that often occured once the short-term stress resolved. Then April 24, 2025 handed us the first case where it blocked a genuinely good trade, the 6.8% cheaper entry from the first half. Weighed against those few barely positive saves, that one real cost settled the balance in retrospect: the insurance premium was no longer worth the payout, and the veto came out.

A few removals we did were of this type: a condition that in retrospect did not offer any meaningful added protection nor helped generate more returns. There was one removal that, with the help of our AI agent, proved that it was pratically impossible to trigger in any realistic ways. We thus decided to remove it changing not much as it had no incident in the past keeping everything tidy.

The results

And the results of all this? The part of the strategy that determines when we enter bear mode, the -20% threshold, did not change. The strategy steps aside on exactly the same days as it always did. Every improvement came from what happens afterward, particularly the re-entries, with the updated strategy getting back into the market earlier in six of the thirteen bear markets since 1957. A few rules that could not justify themselves were also removed, which is a win on its own: fewer rules means fewer degrees of freedom and therefore less opportunity to fit historical luck.

The impact of the changes reduces in modern history as other metrics often triggers a WU In on a bounce that eventually becomes the final Bear End rally. Earlier Bear Market did not benefit from those metrics thus have more space for improvements. Yet, seeing the 18 days gain in 2025 is a good sign that the changes were justified across the history and were not simply tuned to early past occurrences.

In terms of performance, we report our results starting from the end of the previous bear market in mid-2003. Over that period, the annualized return increases from 12.4% to 12.6%, compared with 9.4% for simply holding the market through everything. Keep in mind that we are not talking about the performance of the full hedge strategy here, but only about the impact of the bear mode: we simply hold the market until it officially enters bear territory, at which point the bear strategy takes over already minus 20% down, before fully re-entering when it signals that the bear market is over. More importantly, across the entire backtest, the strategy never incorrectly called the end of a bear market only to be followed by another significant leg down — precisely the mistake an end-of-bear signal cannot afford to make.

We chose 2003 as the starting point because it aligns with the availability of several datasets used by the complete bear strategy. We can—and did—test much of the strategy further back, all the way to 1957, but some of the datasets used to navigate the moves inside a bear market simply did not exist then, so the earlier backtest runs with fewer tools and naturally does not perform quite as well: over the full record the same comparison reads 8.4% to 8.7% against 7.8%. As for how the numbers are built, the block is long by default and only bear mode ever changes that, so outside bear markets the strategy and buy-and-hold are identical, and the whole gap is earned inside the thirteen bear windows.

The number we watched more closely is risk while the block is in charge. Measured from the step-aside point, the worst the strategy ever sat below it during a bear is -10.7% (the 1987 episode, buying bounces into the December retest), the modern worst is -7.1% (the March 2020 bounce, entered three sessions before the exact bottom), and the median episode dips barely -2%. The bounce trades themselves offered 17 round trips opened and closed inside bear markets since 1957, 59% winners, and the worst single trade in 69 years lost 4.3%. There is no disaster trade in the list, which is the entry safety bar doing its job. The failed ones simply caught a real wave that did not generate sufficient profit before reversing to be able to exit early enough to lock gains.

The most complete way to see the block's work is per bear: while the bear block was in charge, what did the market do and what did we do? Being precise about that window matters, so here is exactly how each row is computed. The clock starts at the minus 20% trigger, the moment the bear block takes control, and stops at the bear-end call. Everything the block did inside that window, the step-aside, the bounce trades, the re-entry, is compounded into one number and compared against simply holding the market over the same stretch:

One thing to keep in mind from this definition, which matters for reading the table honestly. The bear-start call comes already -20% from the previous peak. This combined with the fact that the bear-end call only comes once the recovery is underway means that mild bears will show the market up double digits "during" a bear.

With the measuring stick clear, the thirteen bears since 1957, showing the original bear block and the reworked one side by side. The rows are grouped by how much of today's toolkit existed at the time, which makes the leaner-version caveat from above visible instead of abstract. One detail for fairness: since the two versions can call the bear end on slightly different dates, each row grades both over the same window, the longer of the two, so improvements are visible instead of hidden inside different end dates.

Price rules alone (most of today's inputs did not exist yet)

Bear episode	Market	Original block	Reworked block
1957-58	+21.2%	0.0%	+5.4%
1962	+15.5%	+1.8%	+4.9%
1966-67	+20.9%	0.0%	+4.7%
1970	+0.9%	+11.8%	+11.8%
1973-75	-18.9%	-7.6%	-5.9%
1982	+17.2%	+10.9%	+10.9%
1987-88	+22.5%	-0.7%	-0.7%

Part of the toolkit live

Bear episode	Market	Original block	Reworked block
2001-03	-19.2%	+12.6%	+12.6%

Full toolkit live

Bear episode	Market	Original block	Reworked block
2008-09	-35.4%	+16.2%	+16.2%
2018-19	+16.0%	+11.8%	+11.8%
2020	+15.9%	+19.9%	+19.9%
2022-23	+5.4%	+26.8%	+25.5%
2025	+15.4%	0.0%	+6.6%

The rows that moved are exactly the earlier re-entries listed above. The one row that ticks down, 2022-23, is the honest price of buying earlier: the reworked version re-entered six days sooner, just before a brief dip that recovered within days making the earlier price slightly higher than the one from the original bear mode. And notice the pattern across the three groups: the fuller the toolkit, the better the block handles its bears. The price-only era is a mix of solid saves and small losses, while every bear fought with the full toolkit ends positive, which is exactly what the added tools are for.

The softer rows need their context. Every window here opens at minus 20%, because that is the only point where the bear block alone can act. In the modern era, the full Hedge Signal was out before that line printed everytime, from nine sessions early in 2020 to five months early in 2001. The 2025 row is the concrete case: the table grades the block at +6.6% against the market's +15.4%, while in reality the full signal stepped aside on December 18, 2024, seventy-two sessions before the minus 20% headline, and is back in on April 24 below where it left:

It's always a good idea to look at each entry since its possible with only 13 events. However, having an aggregate metric is what help us see how the mode stands across the full history. One standard way to do it would be to run it as a strategy compared to buy and hold from the end of the previous bear in June 2003 to now. Doing this would say that we accumulated roughly +1450% over 23 years compared to +690% doing buy and hold. I, however, do not like this approach as early compounding can hide later neutral to slightly bad performance. If a strategy lock lucky gains early, the head start will compound even if the strategy does not overperform buy and hold for the rest of the window.

This is why I often prefer to show aggregate stats as the average/median/max/min of the thirteen episodes which avoids letting one lucky early episode inflate a compounded total. Per bear, the market did in average +6.0%, the original block +8.0%, and the reworked block +9.5%. In the three bears where the market actually lost money, it averaged minus 24.5% while the reworked block averaged +7.6%.

Wrapping up

The net result is a bear block with fewer parameters that are all cohesive, a slightly better historical record, and the same robustness. Fewer parameters matters beyond tidiness as a strategy with less to tune offers less room to fit noise in the historical record.

Our next job going forward is reviewing the bull market side of the hedge signal doing the same kind of rework we just presented. This will be the subject of our next blog update in a few weeks time. We will then do a final round about incorporating new datasets and ideas we've had to see if there's anyway to improve it further. You will hear more from us in the coming weeks/months on this process.