Bernstein was a neurologist who got obsessed with a T. Rowe Price study showing a four-asset portfolio beating 75% of professional money managers. So he built his own spreadsheets, ran the numbers back to 1926. What he found is the reason "robustness beats performance" isn't just something I say, it's something the data has been screaming since before most of us were trading. The core move of the book: take five different historical windows, calculate the mathematically perfect allocation for each one with full hindsight, then watch every single one of those "optimal" portfolios get destroyed the moment you carry it forward. 100% Japan in 1989. 99.8% precious metals in 1970. Every one a genius pick looking backward, every one a disaster looking forward. That's what happens when a fancy-looking tool gets fed recency bias.a

Okay so I finally sat down and chewed through Ilmanen's Expected Returns front to back, all 1000+ pages, not the highlight reel. Figured I'd write up what actually matters for us, because there's one idea in there that's basically the whole bootcamp philosophy validated by a guy who used to run this stuff at Brevan Howard. Here's the thing that jumped out. Equity premium, credit spread, currency carry, short vol, illiquidity premium, momentum's crash-hedge behavior, on the surface these look like twelve unrelated trades. They're not. Strip away the packaging and almost every single one of them has the exact same shape: small steady wins, funded by a standing exposure to rare, correlated, badly-timed losses. You're not finding twelve edges. You're finding one edge wearing twelve costumes, and the costume is "sell insurance and collect the premium until the hurricane shows up." That's the casino factor, dressed up in academic language. Nothing new to you guys if you've been through the lecture, but it's wild to see it independently derived across an entire career's worth of institutional research. The part that should actually change how you think about your backtests: risk isn't about how big the drawdown could be, it's about WHEN it happens. An asset that loses money specifically when everything else is also losing money gets punished way harder than raw volatility would suggest, because that's the exact moment you have zero flexibility to wait it out. Same reason I keep hammering regime awareness over raw Sharpe, a strategy that's flat-to-up during 2020-style shocks is worth more than a strategy with a higher standalone Sharpe that happens to blow up in sync with everything else. You're not being paid for volatility. You're being paid for surviving correlation spikes with your hands still on the wheel. Second thing, and this one's a gut check for anyone who's found something that works great in backtest, Ilmanen's got a whole section on how a real risk premium can curdle into a crowded,

Here's the part that ties all of it together, and it's kind of a funny one because it sounds backwards until you actually sit with it. The people chasing higher degrees of freedom think they're building something more advanced, more evolved, closer to actually seeing the machinery underneath the market. But what they're actually building is something with a shorter shelf life, and I mean that almost literally — the more finely you fit a model to a specific stretch of data, the more that model is a photograph of a moment that's already gone. Regimes change. Correlations flip. Volatility clusters differently every cycle. If your edge only shows up when fifteen conditions line up exactly the way they did in your training window, you haven't found a law of the market, you've found a fingerprint of 2021, and fingerprints don't repeat. Compare that to momentum. It's dumb. It's been dumb since before any of us were born — it shows up in equity data going back over a hundred years, across different countries, different market structures, different eras of technology, different central bank regimes, wars, booms, everything. It didn't survive that long because someone fit it perfectly to any one of those environments. It survived because it never tried to be precise about any of them. It just says: things that are going up tend to keep going up for a while, and things that are going down tend to keep going down for a while, because humans are the ones trading and humans herd. That's it. That's the whole insight, and it's basically insulting how simple it is — which is exactly why nobody wants to believe it's the answer. It doesn't feel like you earned anything. There's no clever math flex, no PhD flex, nothing to post about that makes you sound like you cracked the code. It's a Toyota from 1995. It just runs. And the reason the fancier stuff feels so tempting is that it feels like power. People want the nuclear weapon, right, they want to feel like they're the ones who finally figured out how to read the tape underneath the tape. But a random system doesn't care how much firepower you point at it — it just means you now have more ways to be wrong at once. Every extra decision layer isn't an extra unit of intelligence, it's an extra roll of dice stacked on top of the last roll. A coin flip has two outcomes. Stack five conditional coin flips on top of each other trying to "confirm" a signal and you haven't built certainty, you've built a slot machine with five reels instead of one, and now you need all five to line up instead of one, which means you trade less, and the trades you do take are the ones that already happened to look perfect in hindsight during backtesting — survivorship dressed up as sophistication.

Each book covers a different failure mode. Carver gives you the operational discipline to actually run a system instead of just theorizing about one. Masters stops you from fooling yourself with overfit indicators before you even get to ML. López de Prado stops you from fooling yourself with overfit ML. Ilmanen stops you from expecting returns that don't exist in the data. And Wilmott sits underneath all of it as the theory backstop — reminding you that every model is an approximation, and the moment you forget that is the moment you blow up. The Reading List 1. Advances in Financial Machine Learning — Marcos López de Prado The book that put "your backtest is probably lying to you" into a formal, citable framework. Covers meta-labeling, purged cross-validation, fractional differentiation, and why most published trading strategies don't survive out of sample. Dense, math-heavy, and worth every hour. 2. Expected Returns — Antti Ilmanen Not a trading book — an asset-return book. A brutally thorough survey of what actually drives long-run returns across equities, bonds, credit, and alternatives, and why most investor intuition about "risk premia" is wrong or overstated. This is the book that keeps you honest about what edge is even possible. 3. Systematic Trading — Robert Carver The most practical entry on this list. Carver takes what he ran at AHL and turns it into an actual playbook: position sizing, instrument diversification, forecast scaling, and how to build a rules-based system you can run without flinching. Less theory, more "here's how you actually do it." 4. Statistically Sound Indicators for Financial Market Prediction — Timothy Masters The book most people skip and shouldn't. Before you build any model, you need indicators that carry real information — not noise dressed up as signal. Masters (ex-spy-photo analyst turned quant statistician) gives you the tools to test whether an indicator's edge is real or just overfit luck. 5. Frequently Asked Questions in Quantitative Finance — Paul Wilmott

Dear friend, Do you want to know why some trends pay out for two years straight... ...while others die two weeks after you finally got in? (This is a free skool where i'm posting pieces that educate like this all the time for free so join as a member if you haven't already) The answer is not luck. And it's not "the market makers hunting your stops." The answer is that a trend has an anatomy. And once you can read it, you'll never look at a chart the same way again. So grab a coffee — or a White monster if you're a real one — and read this slowly. Before you can trade trends, you have to understand what actually causes them. Because a trend is not caused by one thing. It has an anatomy. And the anatomy changes depending on how long the move lasts — which is exactly what the academic triple momentum literature figured out with their 1, 3, 6 and 12 month windows. Different horizons don't just measure different speeds. They measure different CAUSES. Short-term momentum moves off temporary stuff. An earnings beat. A headline. A Fed soundbite. A squeeze where every short has to cover at once. The herd reacts in days — FOMO on the way up, fear on the way down — and then the fuel runs out, because the cause itself was temporary. The headline gets digested. The shorts are done covering. There's nothing left pushing. That's why the research finds short-term momentum the most prone to reversal: the move outruns its reason. Medium-term momentum is a different animal. That one moves off extended sentiment — a story the crowd needs months to digest. Think of a company raising guidance quarter after quarter. Or the AI datacenter buildout, where every month another capex announcement stacks on the last one. People underreact at first. They anchor to the old price. Then they slowly reprice as the evidence keeps coming in. Jegadeesh and Titman found this is the most consistent zone of momentum, and the reason is simple: the causer itself persists for months, so the move persists for months.