Not Losing Money and How CAPY Works

My investing philosophy and the system behind it

Rather than leaping toward macro predictions or portfolio companies, or my AI valuation system, I want to start at the opposite end and think about what’s most important (to me at least), which is not losing money.

Risk

To understand my “not losing money” strategy it’s worth comparing any plausible alternative to the marginal strategy. We’re blessed in 2026 to have a fantastically effective, tax advantaged, and trivially executable investment strategy at our fingertips – buying the index. Whether it’s SPY, VT, some combination or what have you, it’s possible to get <0.2% expense ratio, extremely diversified indices held by reliable counterparties like Vanguard or BlackRock. I suggest these are fantastic investment opportunities and folks forget about them because they’re boring. What’s more, at this point getting a real return of 3-5% is plenty sufficient to maintain my lifestyle, so there’s enormous negative utility in trying to do something clever and blowing myself up. The only reason I try is some combination of being AI-curious and trying to beat the game, for better or worse.

If we buy say VT (Vanguard Total World Stock ETF) what do we get? Well, we get a portfolio of mega- through small-cap global securities, market cap weighted, with an expense ratio of 0.06%. What does that mean, in real terms? It means we lock in the expected return of global public markets, in perpetuity (no need to rebalance ever), and we need only pay tax on dividends rather than on capital gains. We can confidently hold this until we’re dead. That’s perhaps a 7% real return, with max drawdowns of perhaps 50% in our lifetimes peak to trough (could be higher) and almost no real risk of ruin. Note that if we simply duplicated the returns of this basket but have to shuffle things around, pay transaction fees / slippage, and cap gains rates of ~25% then our necessary expected real return simply to beat this hurdle rises to close to 9.5% annually, or 12.5%-ish nominally. The fraction of active managers who meet or exceed this threshold over 10Y periods is very low, it depends how you count, but no more than 20%.

Index hurdle: the passive, low-tax alternative is the real benchmark

OK, so is the bar high enough yet (we must be in the top ~quintile of active manager performance to beat the VT+sleep strategy)? I don’t think so. First, many active managers smuggle risk in from varying angles. The main three I observe are (1) concentration, (2) beta hacking, (3) factor stacking.

Concentration is simple; it’s just owning a small number of securities. You can do a thought experiment here. Say I only owned one stock, call it AAPL. AAPL has idiosyncratic vol (i.e. non-beta vol — e.g., iPhone sales decline because they misexecute), both to the upside and to the downside. The idiosyncratic stuff tends to have momentum, i.e., if I’m right about the trajectory of iPhone sales I’ll be right a few years in a row as I’ve correctly predicted a long duration causal variable bearing alpha. Thus, if I only own AAPL, I can look good for a fairly long time if I’m right. Then when I’m wrong, it all reverses.

Many investors own 5-15 stocks, and many of their causally relevant theses affect multiple stocks within their holdings. Thus they’ll be right for a while, but when they’re wrong things will reverse. The time duration of this effect can be quite long, 10+ years in some cases. Thus some of the 20% of outperforming investors are just happening to run well with a concentrated book, even assuming there was also alpha. I suggest this is common. But it means ceteris paribus you should adjust track records towards the mean quite strongly absent some observation that’s strongly countervailing the concentration effect. It’s also worth mentioning concentration is unavoidable for the investor; we can only know so many companies well.

Moving to beta hacking. Here we take on more risk and mistake it for return. The index has beta =1.0. If our portfolio has beta >1.0, then our benchmark should be the leveraged index (less the other negatives discussed, namely fees, taxes and concentration). Many managers run high betas, then sell their funds based on the “alpha”, which is really just leveraged beta in disguise. A classic example of this would be the ARKK innovation ETF, with a ~1.6x beta. Since 2014, it has (a) actually underperformed SPY (~12.8% CAGR vs ~13.6% for SPY since 10/31/2014 inception), (b) had dramatically higher beta, and (c) charged more fees.

If you’ll let me get a little technical, it’s worth mentioning there are numerous and more insidious versions of beta hacking. Many securities’ betas are calculated in disparate ways, e.g., in Bloomberg, and without getting too far into the weeds I’d just underscore that smaller companies’ betas cannot be trusted and should be computed manually. Even larger ones contain error. The result is that many concentrated investors comfort themselves with a manageable sub-1.0 beta on their terminal while their true exposure to market fluctuations might be 1.5x or 2x their quotation. This usually only becomes apparent during a negative liquidity event or recession, at which time their purported alpha blows up.

Finally, factor stacking. Many investors take on enormous factor betas in addition to the concentration and beta stacking we just discussed. A factor is a portfolio’s sensitivity to a specific market segment. Traditionally that would be, say, value vs. growth, with the idea being value tilts carried alpha historically. However, once folks know a factor carries alpha, it gets arbitraged away (or often it contains a hidden risk premium that cannot be arbitraged, e.g., liquidity premia). The logical conclusion of this is factor explosion, as more and more finer factors are identified, back-tested, and gambled on to find the next big thing. Today you can find a Goldman Sachs factor for AI, for AI memory, for optics, or most anything else. And of course as fast as they can be discovered you can be sure they get arbitraged away near as fast, so it’s a shell game and of course Goldman is always winning.

Now, being long a factor is not inherently bad, in fact all concentrated investors and most active investors are long multiple factors and sub-factors, knowingly or unknowingly. This is sort of obvious — if you’re long NVDA, you’re probably long AI factor, and non-ASIC share factor, or god knows what others as part of your thesis. The problem becomes when so much of your book is levered to a particular thesis or factor that being wrong becomes catastrophic. For most investors, 4-6 factors dominate their portfolios, which isn’t so bad if these don’t interact. But look at other investors and you see for example that everything is levered to AI, and so if that reverses, a 20% drawdown in SPY might be a 60% drawdown for their portfolio. This is both hard to handle emotionally and quite bad for long-term compounding math (aka the elusive free lunch).

Note the rhyme between all 3 of these properties: concentration, beta hacking and factor stacking. All active investors concentrate which masks their track records, provokes mean regression in performance and raises volatility. Most managers beta hack and hold betas >1.0 when properly considered. Nearly all factor stack, some egregiously so. Taken together, this means most active managers have far more volatility, far less alpha, and far more blow-up risk than they’d like to admit.

Risk smuggling: concentration, beta hacking, and factor stacking can all masquerade as alpha.

Where does this get me? My view is <10%, and perhaps <5%, of active managers are actually beating the market all things considered. My bar is two-fold: (1) I want to be taking less risk than the index, and (2) I want a higher after-tax return than the index. I just think you must be very good to do this consistently. And that’s before charging any fees, paying myself for my time, research/compute costs, etc.

The fact I bother is possibly a mild form of insanity.

Strategy

My general strategy is almost embarrassingly simple: I’m a value investor. What does “value investor” mean? For me, it’s figuring out what stocks are worth, then buying them for substantially less than that. This means I need three quantities: (1) the market price, (2) my price — expected intrinsic value per share or E[IVPS], and (3) a bridge function from market price to my price, which is the expected internal rate of return or e[IRR]. I’m just as happy buying a sexy pre-revenue tech startup as a boring industrial conglomerate so long as the numbers are good. I always think about risk, so if E[IVPS] is 2x market price but my adjusted beta is 1.5x, I’d usually consider that a worse opportunity than a stock with 1.75x market E[IVPS] but beta 1x. You can think of this as an adjusted expected return and as a means by which I optimize my risk budget.

Value bridge: E[IVPS] estimates value; E[IRR] asks how and when value becomes return.

Why not other more complex frameworks? Certainly I can understand math, factors, etc. The problem I see is these frameworks represent a statistical layer superimposed on the causal variables driving business performance. In other words, I might write a clever algorithm or AI system to trade in a way that makes alpha based on correlations between synthetic variables. The problem I see here though is that at the end of the day I’m now playing a higher-dimensional numbers game against other sophisticated players. There’s a bit of a moral argument here too; when I buy a stock it’s because I think it’ll actually make the stuff it says it’s going to make and folks will want that stuff. In the high-dimensional shell game you simply optimize for price movement. When wrong, it tends to fall apart.

In contrast, I try to find out what’s causal via fundamental analysis. The idea is that at the end of the causal stack, what’s driving the performance of AAPL stock is how AAPL is run, how it’s positioned, and how good they are at making products people will want to buy. In other words, are they a good company in the classical economic sense? If yes, I expect the stock to go up over time. If no, I don’t. Simple conceptually but difficult practically. Less downside when wrong, and a slower pace.

A Bit of History

At a certain point in my career, I had enough money that I could start investing in a potentially more interesting way, but a 7% real return in the index didn’t seem that exciting.¹ So I read about many strategies and for the reasons above settled on the tools of fundamental analysis. I figured that if owning 20 random companies gets you ~95% of the index’s diversification at the same expected return, then even if I was bad at investing I’d lose only some of my money rather than all of it.

I began by what I call “discovering the past.” In other words, I figured the lowest-hanging fruit possible were securities where all I needed to do was read the balance sheet, find a sleepy operating biz, make sure management weren’t actively defrauding me, and do simple math. Now, it’s never that simple unfortunately. But the idea was to buy $2 of bananas for $1 and let the chips fall where they may. This went OK; I didn’t make a ton of money but I didn’t lose my shirt either and I learned along the way.

Over time I tried to broaden my approach to “understand the present.” This means valuing operating businesses much more granularly while paying less attention to fixed assets like tangible book value as downside protection. There’s more risk and reward here. A cyclical trading at 5x FCF is a trap, while a quality biz at 20x FCF might be a bargain. Here I’d attempt to understand and model businesses using non-heroic assumptions and find ones that are undervalued. To a large degree I’d say I’m still here. Note that the past, present and future all blend into each other in securities analysis. One minute you think you’re discovering the past and you realize IFRS book value contains assumptions about the future. Thus there’s no real epistemic safety but there are certainly higher or lower difficulty settings and associated rewards.

The final hurdle is “predicting the future.” Here, one articulates an honest base, bear and bull case for a company and can reliably price each case, assign probabilities, and where possible correctly predict bullish outcomes with enough reliability that starting E[IVPS] (aka the objective value of the present under reasonable assumptions) no longer becomes the guiding epistemic guardrail. This is the pinnacle of investing and even the GOAT investors practicing this style, (e.g. Stanley Druckenmiller) suggest they’re right only about 40% of the time, and also rely on valuation as well as other tools to limit their downside. In other words, predicting the future is really hard. Ironically most bad investors spend most of their time trying to predict the future. At present I do put my finger on the scale and make predictions about the future (both positively and negatively), but I tend to stay very grounded in valuation in order to avoid having to be right very often. I prefer situations where I can be wrong frequently and still do well.

Difficulty settings: more future underwriting can mean more reward, but less epistemic safety.

Making Things Easy

If you’ll permit me a couple of poker analogies. In poker, there is GTO (or “game theory optimal”) poker vs exploitative (explo) poker. In GTO, a fairly large subset of your opponents’ deviations from the optimal strategy lead to positive expectation for you if you play GTO and pay no attention to what your opponent is doing. In other words, if you play technically superior you’ll get free expectation and can never shift to a negative expectation regime. In explo poker, you make assumptions about your opponent’s play and deviate from GTO to make extra money. If you’re right, you’ll make extra and if you’re wrong you’ll lose extra. All great players do both.

To me E[IVPS] is akin the GTO value and makes no reference whatsoever to the market price. The e[IRR] bridge is a concession to actually getting paid, and me putting my fingers on the model to predict the future. But broadly speaking, I adhere quite close to GTO with the view being I win in both cases either (a) market realizes the discount to intrinsic value or (b) the biz improves and I get paid in the form of dividends/buybacks. This means I usually don’t buy things just on vibes, narratives, good-looking numbers, multiple expansion and so on.

The second poker analogy is on game selection. In poker, there can be a brutally negative expected return game. At a 6-handed table, it’s common for rake to be on the order of a -4% expected return per player per hour. Overcoming a negative expected value like this means that players have to be extremely skilled, but more importantly, they have to almost always play with at least one extremely bad player (or a few pretty bad players) to make money. There’s an obvious conclusion: even an extremely skilled poker player has to play in the right games most of the time, or else some combination of rake and variance is likely to kill their expectation. The problem becomes that as a skilled poker player moves up the stakes, the setup at a 6-handed table is almost always the same: 5 excellent to world-class players and 1 very bad player. In order to use their bankroll effectively, they must play high stakes, and when they play high stakes, there will be a combination of high rake and brutal competition pushing their edge down.

A few poker players manage to escape this dynamic. The classic way to escape this is to be a bum hunter (a heads-up player who only plays very bad players). Bad players are pretty easy to identify heads-up quickly for a skilled professional, because they make very predictable statistical mistakes that you can easily exploit. A heads-up bum hunter often will win closer to 20% per hour with relatively low variance against only weaker players. A player like this is essentially playing in the softest games possible whenever they can. They have full control over who they play against, because the ratio of good player to bad player is always 1:1, and they can quit at any time if they think things change.

Although there’s no relevant analogy for bum hunting in investing, per se, the analogy that extends reasonably well is that most investors choose to play in relatively tough games when I don’t believe they have to. What makes a tough game in investing? Usually, it’s the market cap of the company, expressed in terms of the amount of free float, as well as the investor base and jurisdiction of the company.

To play in the softest investor games, you want to be in small, out-of-the-way companies that professional funds and sophisticated investors have overlooked. Of course there are exceptions. You might have a thesis — say that chip design is a great industry — that you can only express through a large- or mega-cap. Take TSMC: it’s an incredible monopoly, and there’s no way to get TSMC-like exposure aside from buying it, because there’s only one TSMC. Its size is exactly what makes it unassailable. So I don’t want to overemphasize small companies to the exclusion of everything else.

But on average, the insight is this: like the poker bum hunter whose edge comes from only playing against weaker competition, an investor not managing billions can get something similar by mostly investing in stocks with a market cap of $1 billion or less. There’s little sell-side research, little public information, and few funds participating. The marginal price is being set by relatively unsophisticated retail investors, not the big shops with proprietary models, dedicated analyst teams, huge research budgets, and serious compute. This has been a pillar of my strategy for a long time now: mostly participating in companies of very small size where I feel my analytical edge is largest relative to the competition. Investing, I say, is hard enough already and I don’t need to make it harder. Plus I figure that to the degree I’m doing of marginal social value, it’s better that I’m setting supply/demand for some obscure European micro cap than being a drop in the ocean of NVDA.

There’s a downside to out-of-the-way investing. Liquidity is thin, there’s more fraud, disclosure is worse, governance is worse, there’s more insider trading, it’s often in out-of-the-way jurisdictions and risk is higher in many other idiosyncratic ways.

There’s a corollary to this that’s important: many of my strategies that I feel are alpha-bearing at the nano-cap through mid-cap level are no longer alpha-bearing when you look at the large and mega-cap level, where the quality of analysis is dramatically better. In other words, I suspect that I am creating alpha by valuing a random security in the Russell 3000 or a small-cap security in Europe, but I would guess that I am adding very little, or perhaps even negative expectation, when trying to value a security in the S&P 500.

Soft games: analytical edge is largest where attention and institutional competition are weakest.

Over time, I’ve branched out a bit, and in general I attempt to play relatively close to optimal strategy. In other words, I pay relatively little attention to what I think other market participants are going to do and why. I mostly try to invest in small, out-of-the-way securities where I feel the pricing is structurally weaker.

A Quiet Epiphany

If you read the genre of investment letters, pitches, analysis and so forth, I think with reflection it becomes apparent that there’s a bit of frenzied cognitive dissonance to the commentary. On the one hand, investors know markets are risky. On the other hand, they feverishly pitch their latest idea and “underwrite the risks” by making arguments that if taken literally imply the risk is close to nil or at least closely bounded. I can relate emotionally to this. In my first years of investing, it was like each new portfolio company I researched became more and more attractive to my mind, until before final investment I could barely imagine ways it could go wrong and wanted a 12% position. And yet, something like 10% of ideas I invest in get distressed or go to near zero, many languish, and only a few go great. So for myself and many, it’s extremely clear that recency bias, proximity bias, availability bias and general optimism bias infect security evaluation.

In post-mortems, often we get the cousins of these biases disguised as learning. Confirmation bias, hindsight bias, and a kind of self-flagellation bias. The investor explains why their thesis was wrong, where they should have cut their losses, and so on. Risks that seemed peripheral or swanlike on the way in materialized. The investor notes down their “lessons” and soldiers onward, better prepared to be the next Buffett. In fact, in many ways I think Buffett and his acolytes are the cause of this hindsight bias. The idea somehow is that the investor can know with something approaching certainty the attractor state of the company if they’re somehow smart enough. Maybe. I haven’t figured it out yet.

Over time, as I got my head kicked in on more and more investments where my rosy perceptions differed from reality, I realized that this idea of knowing attractor states just isn’t realistic. We can certainly model and guess, but reality is too sneaky to pin down. I’ve since come to internalize the idea that buying shares is a lottery ticket into the set of possible futures for a given company. Empirically we can see that for, say, small caps, the standard deviation in one-year share price is on the order of 30-40%. The best we can hope for are asymmetric distributions. It’s only on the portfolio level that anything approaching sanity can occur. Luckily, with enough diversification the insanity becomes antifragile. If X and Y both swing wildly with a 30-40% standard deviation, but X is +50% and Y is -50%, the rebalancing towards intrinsic value is wildly accretive so long as the long-run attractor states’ expectation trends even moderately towards my model.

It’s one thing to intellectualize this, and another thing to embody it. I don’t have any idea where the share price is going in the short run, and I see real risks in the medium and long run. I don’t love any of my stocks, I don’t hate any of them.

Perhaps the other epiphany of using CAPY these last few months is that of my “legacy” (i.e. non-CAPY) picks, many were good, but a great deal were mediocre or outright poor. Removing the lowest-hanging fruit (simply wrong picks, bad reasoning) seems like a much more important driver of expected returns rather than making great picks.

Aha! July 2024

I had one of those rare aha moments while sitting in my Airbnb, at an eco-tourism stay in Tuscany in July 2024, while I was traveling with my family. I remember being very stressed as I was behind the wheel, driving through the old, twisting Italian villages in Tuscany and trying not to get my family in a car accident or go the wrong way down one of these medieval streets.

I was relaxing in the easy chair in the middle of the living room back at the cottage, reading on my phone to relax and use my mind a bit, when I stumbled on a pretty amazing article (arxiv.org/abs/2407.17866).² (The paper was later withdrawn, so I treat it as the spark, not the evidence.)

I’ll spare you the details, but the takeaway was that the authors used GPT-4 to predict the direction of earnings for stocks. Crucially, they noticed that even these coarse LLMs, using only a small amount of the available fundamental analysis data (the income statement and balance sheet), were getting predictions approximately as strong as those from academic machine learning (meaning traditional neural net systems) and those of analysts. At this point, I was already interested in AI and was using GPT-4 to help me in my research process, essentially as generative search. The article persuaded me that it was already practicable to use large language models for valuation in a way that was likely to have some alpha now and potentially significant amounts of alpha in the fairly near future as capabilities improved rapidly.

To that end, I set some guardrails around my project: the entire process would be automated as much as possible, with the input being a security ticker and the output being a structured valuation report, scenario distribution, audit trail, and configurable model. I also decided that each security would be analyzed the same way to make comparison more reliable.

Now, of course, there’s going to be some dynamism in how each security is looked at, but I wouldn’t be building multiple different agentic pipelines to look at varying securities. I felt that this was in keeping with the Bitter Lesson.³ Put optimistically, the view is that if fairly weak LLMs were making alpha-bearing predictions using anonymized financial statements, then this paradigm was highly extensible to stronger LLMs making predictions using the entire document corpus for these companies, as well as anything else they could find on the internet.

This set in motion my project of the last couple of years, which was to turn this germ of an idea into an actual platform that could reliably execute on it. I believe that’s just coming to fruition now. The securities in the present portfolio are primarily chosen by AI and discovered by AI.⁴ There is a human-in-the-loop layer to sanity-check the outputs, but the large majority of the analysis is now going through my standardized process and my AI system, which I call CAPY.

Current Approach

At the present moment, I have a hybrid human/LLM system for discovery (‘discovery’ here means finding new companies to run through the agentic valuation system, CAPY), and then the subset of CAPY runs that have E[IVPS]>= to market price I look at in collaboration with LLMs to do a “human-in-the-loop” audit of the automated CAPY result. I then adjust the CAPY value as needed (or sometimes not needed, the system got it right!). I have a dashboard of values from all securities I’ve observed in a “big board” and then I make allocation decisions collaborating with AI tools.

I find for my very best ideas I suspect E[IVPS] is ~3x market price, while for most of my buys I get ~2x market price. I expect IRRs in the range of 20-40% per model, but realized real-world IRR is likely to be dramatically lower because of ex-model drag — variables unobservable to my framework but partially priced in by the market. Nonetheless I do think it’s possible to exceed the risk-adjusted return of the index using this strategy, and I’m optimistic the current CAPY agentic system is highly extensible to capability improvements at the model and harness level. (I guess I should note most of my stuff runs on Claude Code Opus 4.8 at present; though GPT 5.5 pro + xhigh is arguably superior as of writing. I use both of these models and am model agnostic over the longer term.)

As for sizing, I aim for ~20-30 positions ranging from ~2% of NAV at the low end to 10% at the high end. Some of this is pragmatic. I’m managing enough money that, for certain very small securities, if I were to hit my risk tolerance, it might be an appreciable percentage of the float and might make liquidity constraints bite. Sometimes it has to do with the risk distribution or skew of the security. Sometimes it’ll have to do with the expected return. Often it’ll have to do with all of these taken together.

More practically speaking, to have a say 8% to 10% position in a security, it will often be relatively larger-cap, relatively liquid, well-governed, and have a tighter distribution of outcomes, or at least one with not very much left skew. A 2% position might often be illiquid, a bit opaque, and have lots of skew in both directions. I do think I run a fairly large amount of concentration risk, likely a fair amount of factor stacking risk, and a bit of beta stacking as well — by which I mean my portfolio’s true beta ends up above 1.0, not that I’m ripping beta and selling it as alpha à la the beta hacking discussion above. I’ll go through these in turn and explain how I deal with each compared to the average practitioner.

As I said previously, the classic wisdom is that if you own 20 random securities, you get approximately 95% of the diversification of the index, which sounds pretty good. Practically speaking, however, we’re buying mostly medium, small, and micro-cap securities, often global, and these securities are non-random in various important ways. I’ll explain later, when I talk about factor stacking, how they have certain covariances.

For now, I’ll just concern myself with concentration. The idiosyncratic risk of each of these securities tends to be dramatically higher than that of an average S&P 500 security, so I expect them to swing all over the place. Given that each has typically much higher volatility than an S&P 500 security, and even if those volatilities are mostly idiosyncratic, you need a larger sample to get to 95% of the index’s diversification. That’s why I aim for closer to 25-30 positions here. I’ll probably try to increase the number of positions I hold as I get more skilled with AI and can keep track of more positions.

Right now, I’m still somewhat constrained by the degradation of the opportunity set as I try to widen my book. I think once the CAPY system becomes fully automated, with fewer human-in-the-loop decisions, I’ll be more free to invest in a larger number of securities, potentially up to 50, which should reduce concentration risk somewhat dramatically. Thus, I think when assessing the portfolio recommendations I make here, you should realize that concentration risk is still a drag on the compounding rate of the portfolio in a way that I have not fully solved yet.

Moving to beta stacking, without getting into all of the technical details. Many of the securities I invest in — say, a microcap, illiquid Korean stock — might actually have a sub-1.0 beta on traditional algorithms (say, 0.5 or something). I think this can lull the naive investor into a false sense of security, thinking that the beta of the security actually is what it says. That really is just a function of the liquidity of the security, not its true beta. Even in the case that there’s more liquidity, often the betas are still wrong for various other reasons.

So I have my own way of calculating beta, which I won’t share, but which tends to be higher than what the market betas are. The result is that, for most things I own, the betas are somewhere between 1.1 and 1.5, and that’s how I price my risk. If I’m trying to invest all of my money, and almost everything I own has a true beta above 1.0, I either have to raise my benchmark to something like a levered index or find a way to hedge the tails. I’ve chosen the latter. In ordinary markets, this does not bother me much; if the index is moving around 10% or 15%, I expect the alpha to matter more than the extra beta. The problem is stress. In a real drawdown, small-cap correlations rise, liquidity disappears, and the things I own often trade more like one crowded risk bucket than like independent businesses. That means my true downside beta may be materially higher than the quoted beta, and I may be losing fastest at exactly the moment I most want cash to buy. That is the lacuna in the strategy I’m trying to solve.

It gets worse once we get into factor stacking. As I’ve argued, factor stacking and concentration go hand in hand (if you only own 20 things, some of them are likely for the same reasons) and so I’ll often have extra exposure to specific themes in non-random ways. Even worse, all small cap or “out-of-the-way” securities (which I specialize in) tend to have a shared liquidity risk when times are tough. There’s a simple explanation and it’s the mirror image of the reason why they’re “soft games” in the first place. Most of the investor base is retail or smaller funds. They’re often emotional, levered, or both. When the market crashes, they bail or need money or get liquidated. They sell their small and “speculative” securities. This is more than hypothesis, you can see it in large swaths of empirical data, and I’ve seen it practically. During bad times, my small-cap biotechs and my small-cap Japanese industrials’ correlations rise dramatically and they all puke together despite having totally different leverage profiles, management, domicile, biz characteristics and anything else you could think of. There’s no mystery here, it’s simply that the small cap investors are getting margin called.

I can think of at least three ways to rationally deal with this issue. The first would be the Seth Klarman way of using absolutely no leverage: simply holding through downturns and sticking with your guns, hoping that your companies will re-appreciate. This is entirely reasonable, and for many points in my career, it’s exactly what I’ve done. A more conservative approach would be to leave a reasonably large portion of the portfolio in cash. The idea here is that during downturns, you can redeploy the cash to buy discounted securities. Your volatility will also be more in line with the market because you’re less exposed when the inevitable crashes happen. I’ve chosen a bit of a hybrid approach that’s slightly more technical.

As I’ve argued, I don’t much care about one standard deviation-style moves (in other words, moves of 10% or even 15%) in the broader index within a year or so. I care when those moves become 20%, 30%, 40%, or 50% in the index, and I believe my swings will be larger than the indexes on those tails. To hedge specifically those short-duration, violent moves that tend to swamp my alpha and lead to large disconnects from intrinsic value for the securities I tend to hold, my strategy has been to keep a reasonable percentage of NAV (say 2% to as much as 5%, depending on the climate) in medium-dated index puts. I’ll often time these puts to expire at the end of the tax year for simplicity. I’ll often buy a basket, but mostly with strikes >10% below market price.

The idea here is that the “vega efficiency” of this security is quite high.⁵ Without getting into all of the quantitative details, what this means is that I’ll have some out-of-the-money puts of reasonably long duration. I expect that if the market ever goes into a high-volatility downward regime, the value of these securities will increase dramatically. I’ll be able to sell them to buy some of my extremely discounted-intrinsic-value securities. The cost of this is fairly significant, as the expected return of these puts is very low. I’d estimate around -30% annualized. To make up for that, I have to do very well when I’m able to sell them for cash. I believe the expectation math just barely works and these are close to neutral full-cycle expectation for me. More importantly, this allows me to stay invested with close to 100% of my money at all times, hoping rationally that I’m not taking on either more fundamental risk or dramatically more volatility than the index.

Summing up, how this works practically is that I’ll often have many securities at 2.5% through 6% of NAV, a few edging up towards 8%+. The math tends to work cleanly, if I buy a 4% NAV position at ⅓ of E[IVPS] and it goes up 100% while E[IVPS] goes up less (as tends to be the case), I can comfortably hold that 8% position that’s still at a, say, 60% discount to E[IVPS] without breaking my risk tolerance. As it continues to trend towards intrinsic value I’ll keep trimming, adding to either new positions or positions that haven’t been as fortunate and need a top-up.

It’s worth mentioning this process of dynamically adding/subtracting to positions as E[IVPS] vs. market changes we affectionately call “Turtle Creeking”, after the notable hedge fund who systematized this practice and published an influential study that ~half of their 15% annualized return over 20+ years was due to rebalancing rather than stock picking.⁶

In practice the rebalancing edge gets partly eaten by spreads, market impact, and short-term gains tax in the most illiquid names. The conceptual ideal needs to be discounted by what actually trades. We add the wrinkle of owning vega-efficient puts to dampen sub-index drawdowns (meaning fundamental impairment) and volatility (meaning violent price swings) in times of stress. The hedge isn’t perfect — index puts cover broad-market and broad-liquidity crashes, not every form of small-cap-specific stress, and they have a real cost — but they let us stay close to fully invested without the strategy falling apart in a real crash. We’ll likely be wrong, but at least we’re trying to deal with risk rationally.

We aim to hold securities for 12-36 months. Practically speaking it varies. For example, we purchased a security called CINT.ST a few weeks ago and it promptly got a take-private bid at a 33% premium to the prior close, or about 41% to the 30-day VWAP. We’re not special situation investors, so we exited. Another security, Permanent TSB (PTSB) we held for ~5 years before a similar situation occurred. More theoretically, there’s a very large body of research that suggests that outperformance of securities along many vectors (whether it be earnings, return on invested capital, price, and so forth) regresses to the mean relatively rapidly but with some momentum. Therefore, we’d love to hold a security forever, potentially, if it kept increasing in price and IRR was good and our models continued to suggest that it still was dramatically undervalued. Practically speaking, though, what tends to happen is that we find things that are undervalued; over time they trend closer to fair value, and once they reach close to fair value, we exit for a better marginal situation. This tends to take somewhere between one and three years.

We also aim to diversify ideas across factors, geographies, themes, industries, and market caps. The idea is that we can be massively wrong about a causal variable or two and still do OK.

Summing up, the idea is to use the CAPY pipeline to discover, evaluate, and allocate capital. Much is automated, but there are still some human-in-the-loop touch points. We aim for 20-30 names of 2%-10% weights across many diversification axes. We aim for ~full allocation with no leverage, but consider that even a bit too risky and therefore add some OTM puts to hedge tail risk. The hurdle here is a portfolio that’s at least as safe as the index and exceeds its after-tax expected return.

Portfolio architecture: position sizing, diversification, beta discipline, and tail hedging work as one system

Model Strengths and Weaknesses

Without getting too far into the weeds, I’ll disclose a bit about CAPY. CAPY is a DCF + scenario analysis system, with an attached IRR module. It also includes some “nice to have” tools such as a discovery engine, and portfolio management tools but these are bolt-ons for my own convenience.

The process goes as follows: I initiate a single command, e.g., ‘Run X’ and the system researches, analyzes, and models a given security. This takes a large amount of compute and time, roughly ~75m tokens/security and ~5.5 hours wall clock even with parallelization.

The process is meant to satisfy standardization, auditability and dynamism. For example, I could probably hand-guide the LLMs to make more parsimonious models for any particular name. The challenge was building a fully automated agentic system that gets halfway reasonable output values for E[IVPS] and e[IRR] without any human guidance other than a target company.

There are still many obvious blind spots with the model. Very occasionally it will make accounting errors with complex books, or other technical errors. Sometimes it will make a higher-level conceptual error, and fail to appreciate the right framework for company evaluation. Some companies are either ex-model or difficult to model well within the existing framework. Incentives, chaotic capital allocation effects, and governance can be difficult, as can balance sheet heavy firms. Plugging the whole edifice into a broader world model and macro model is also out of scope for now. I believe progress can be made on all of these axes, but it’s worth mentioning it’s not a magic oracle. Research can always be made more dynamic and more rigorous.

Nonetheless, I’m proud of it technically. A system running for >5 hours wall clock with significant parallelization that’s able to reason over a state space this big, produce artifacts, and output final deliverables that are fully configurable and auditable, is an extremely useful tool.

As mentioned previously, I run a HITL or “human-in-the-loop” audit process for all investments. Here I look at the CAPY outputs, do dynamic research, question assumptions, and then eventually re-run the model using corrected inputs. Often, CAPY gets it materially right and I don’t change much. Sometimes I catch obvious errors. More often than not we get into the realm of interpretation.

Lest this seem like more than it is, it’s worth mentioning that most companies I buy (and which CAPY likes) are extremely obvious, and I can tell if the spot is good in the first 10 minutes. The rest is just diligence.

In fact, one way to frame the whole project is that I’m trying to bet on AI without taking on AI beta. I can do this by using AI intelligently for security analysis while not being overly long AI beta myself.

One thing to be clear about: this is CAPY v1. I do not know whether it is calibrated in the only sense that will ultimately matter, which is whether the names it likes go on to beat the benchmark after tax and risk. That test starts now, with this vintage. What I can show today is less grandiose but still useful: how the first production vintage behaves when pointed at the sort of companies I actually care about, and how often the human audit changes the answer.

The first thing to be honest about is the selection bias. I have not run CAPY on 100 random public companies. I have run it on things I owned, things I almost owned, things that looked odd enough to spend five hours of machine time on, and a few things I wanted as calibration examples. You should expect that set to skew cheap. If a model looking at my own hunting ground did not produce E[IVPS] above market more often than not, that would say something bad about either the model or the hunting ground. But the answer is more interesting than “yes.” In the production artifacts I can verify, CAPY has 103 completed company runs and 102 of those have same-currency E[IVPS] and run-date market-price data. Fifty-nine of the 102, or 58%, had E[IVPS] above the market price used in the run. The median premium was 11%; the mean was 36%; and the 10th/90th percentile spread was roughly -62% to +163%. In plain English: lots of maybes, plenty of no’s, and a right tail that is doing most of the work.

Figure 1: Same-currency CAPY runs sorted by E[IVPS] premium versus run-date market price.

The shape of the distribution is more useful than the average. It is not normal, and I would not want it to be. The model is not sampling the market and estimating fair value in some clean academic sense. It is sorting a biased hunt list into bins of usefulness. Across the 102 comparable runs, the bottom decile had median E[IVPS] 78% below market. The top decile had a median premium of 264%. Both tails matter. The top decile is where the work gets interesting. The bottom decile is almost as useful because it tells me which of my priors did not survive contact with ground up CAPY modeling.

The HITL layer often but not always produces significant change. Across the 64 latest HITL-audited names I found, 61 had either a quantified E[IVPS] effect or an explicit no-change conclusion. Thirty-one moved by more than 20%, and they did not all move in one direction: 19 down and 12 up. The downward changes are the errors one would expect from a young automated valuation system: dilution, governance friction, trapped value, realizability problems, the wrong valuation frame, or too much faith in terminal economics. The upward changes are real too: asset floors, missed optionality, overly punitive discount rates, underweighted event paths, and source-document corrections.

Figure 2: Latest HITL audit per ticker — large E[IVPS] changes cut both ways.

So CAPY v1 does not produce “the right number.” A year from now, with a different generation run and a better error history, it may produce a different number from the same source documents. The point is that it produces a number I can audit, a structure I can challenge, and an error log I can compare across companies and over time. The forward-return test is the current vintage.

CAPY Research publishes experimental, model-generated valuation research and general market commentary for informational and educational purposes only. It is not investment advice, not a recommendation to buy or sell any security, and not a price target. Outputs such as e[IVPS] and e[IRR] are experimental estimates produced by a software model and reviewed by a human; they are not forecasts of actual returns and should not be relied upon for any investment decision. The author or affiliated parties may hold positions in securities discussed. Do your own research and consult a licensed financial advisor before investing.

1. With hindsight I probably should have just bought the index and done something easier, but here we are.

2. This paper has since been withdrawn after the authors flagged data inconsistencies when attempting replication. The withdrawal doesn't change anything for CAPY: the paper was a spark that made me ask the question, not the evidence for the answer. CAPY's methodology — full-corpus document ingestion, DCF with explicit scenario analysis, IRR estimation, and human-in-the-loop audit — is fundamentally different from the paper's approach of using anonymized standardized statements to make directional earnings predictions.

3. Rich Sutton's "Bitter Lesson" (2019): general AI methods leveraging compute consistently outperform methods that rely on hand-engineered domain knowledge, given enough scale. Roughly — don't bake too much human cleverness into the system; let the model and the data do the work.

4. You might argue "You decried the higher-dimensional algorithmic shell game before, and now you're ceding your hard-won causality to a high-dimensional LLM layer!" Fair. But quant strategies typically use leverage, depend on statistical regimes, and decay or reverse fast when regimes shift. CAPY is unlevered and grounded in real businesses. If I were merely average at this and held 25 names, my floor would be ~market return. So it's less risky on a variety of axes — epistemic, leverage, etc.

5. Vega measures how much an option's price moves when implied volatility moves. "Vega efficiency" means buying a lot of vega per dollar of put premium. The point isn't the Greek — it's why this matters strategically. If my alpha is decent, I don't really care if the market drops 10% in a year; I'll do fine. What I do care about is the illiquidity tail where my small-caps puke harder than the index when shit hits the fan. Vega-efficient OOTM index puts surge precisely there. When forced sellers are dumping, I sell the puts for cash and become a buyer rather than a forced seller — and I rarely underperform the market in real stress. I also sleep better at night.

6. https://www.turtlecreek.ca/wp-content/uploads/2024/04/Turtle-Creek-2023-Annual-Letter.pdf