Alan Koska’s Post

Every backtest I’ve ever been shown looked good. Nobody brings the bad one to the meeting. So over the years I’ve ended up with a short list of questions I ask before I believe any of them. Which period was held out, and when was that decided? What did a trade actually cost? How many versions did you try before this one? Could anyone rebuild this without you in the room? Twelve questions in total, one per slide, each with the evidence you’d need to check the answer and how bad it is if there isn’t one. The useful part isn’t Pass or Fail. It’s the third answer: “I can’t verify this with what I’ve been given.” Say that out loud in an investment committee and watch what happens to the conversation. I’ve been on both sides of that table. These are the questions I’d want asked of our own work, which is why the whole list is public. The version you can fill in during a meeting is in the comments. Educational material, not investment advice.

Sharpe above 2, a smooth three-year curve — would you allocate? Every position in it was a coin flip. We built a machine that produces impressive backtests out of pure noise: a random-walk market with nothing to find, 5,000 random strategies, pick the best one. In-sample it shows a Sharpe of 2.27 and a 12% drawdown while the market drew down 40%. Then we break the seal on two years it has never seen. Ten slides on how a fake track record is born — and why out-of-sample data is the only witness. The machine runs in your browser — about 100 lines of JavaScript, nothing computed on a server. Link in the first comment. Simulated data · educational demonstration · not investment advice.

On Saturday we re-ran every backtest in our evidence explorer. 602 of them: simple rules on US, UK and European shares, over 5 to 20 years. 79 made more than buying the same shares once and holding them. That is 13%. 445 had a smaller worst fall than holding. By type, the count that beat holding: momentum 43 of 108, rotation 20 of 44, trend-following 2 of 74, breakouts 1 of 48, risk parity 0 of 48. So mostly these rules cut the drawdown and cost you return. Worth knowing before you pick one. Dividends, trading costs and UK stamp duty are in. The share lists are today's, so survivorship flatters every row, the benchmark included. Hypothetical results, not a guide to future returns and not advice. All 602 are public on Quantis Trade, which went live this week.

Quant trading used to need a maths degree. Not anymore. Course 3 lands on YouTube later today: complete quant trading with AI, from zero to a working system. Free. The old barriers were computation and software that costs more than most people's first year of returns. Agents collapse both. The split that matters now: you bring the judgement, they bring the maths. The course shows exactly where that line sits, because putting it in the wrong place is how people lose money. From tonight, search "Complete Quant Trading With AI" on YouTube. Built for a slow weekend watch. Would you trust a system you built but didn't compute yourself? #quanttrading #aiagents #claude #investing

Lab Demo 01 is live. The Overfitting Machine builds a Sharpe-2 backtest out of pure noise and lets you break the out-of-sample seal yourself – in your browser, source auditable, no sign-up. All data simulated - not investment advice

There is a moment in research that feels like winning and is actually a warning. A test comes back and the numbers are gorgeous. Everything lines up. Nothing is ambiguous. The textbooks are blunt about what that usually means. An effect that extreme is rare in a real market, and the more extreme it looks, the more likely the explanation sits somewhere other than the market. Future information that leaked into the test. A cost that was never subtracted. A filter that quietly removed the hard days. A parameter tuned until the past finally agreed. Statistical confidence works against the beautiful result too. Run enough tests and one of them will look brilliant on luck alone. The strength of a single number says very little until you know how many attempts are standing behind it. So the response to a gorgeous result is not speed. It is a slower second pass. Refit on data the idea has never seen. Rebuild the cost assumptions from scratch. Try to break it on purpose, then hand it to somebody whose job is to disagree. An idea that survives that is worth taking further. An idea that only looked beautiful was never an idea. It was a reflection of the data it was fitted to. Beat the market. Back every idea with data.

The Trade-Offs of Classic Algos TWAP and VWAP are foundational tools, and you absolutely need to understand them. But like any tool, they have their strengths and their weaknesses. My personal rule of thumb is to be acutely aware of the trade-offs I'm making. Both TWAP and VWAP are excellent at one thing: minimizing your raw market impact by being 'passive'. They slice your order up and patiently feed it into the market over time. The major downside? **Timing Risk** (or opportunity cost). Let me paint a vivid picture: You need to buy a large block of stock, so you set a one-hour TWAP. But what if that stock is in a powerful, non-stop uptrend for that entire hour? Your TWAP will dutifully keep buying for you... at progressively worse and worse prices. You avoided impact, but the trend killed you. I find these algos work best in range-bound, mean-reverting markets. In strong trends, they can be painful. What's your experience been? Share your stories! To find out more, visit https://quantza.co.uk

A useful idea from intuitionistic logic: Not disproven is not the same as proven. That distinction matters a lot in systematic trading. A strategy surviving a backtest does not prove it has alpha. A risk not materialising does not prove the portfolio is safe. A signal not being wrong does not make it right. Markets operate under incomplete information. Sometimes the correct state is not true or false, but simply not yet established. That is a useful way to think about research, signals, risk, and execution: Only act on what you can actually justify. In systematic trading, insufficient evidence should lead to no trade, not a forced conclusion. #SystematicTrading #QuantResearch #RiskManagement #cpzai CPZ®