On the half-life of a signal
Every signal decays; the only question is the slope. We walk through how we measure factor decay, why it should set your rebalance frequency rather than your gut, and what happens when you ignore the half-life entirely.
Not the strategies — the discipline. These are working notes on the problems that quietly kill quantitative research: signal decay, survivorship, overfitting, costs, and the reproducibility you can't fake.
Every signal decays; the only question is the slope. We walk through how we measure factor decay, why it should set your rebalance frequency rather than your gut, and what happens when you ignore the half-life entirely.
It's a research bias. When a history quietly drops the companies that delisted or failed, every backtest looks better than it was. We explain why point-in-time data is the difference between a strategy and a survivorship artifact.
Multiple testing, parameter search, and the quiet pressure to show a line that goes up. On how we keep our own backtests honest — and the specific traps we built the platform to refuse.
The strategy that looks great before costs and mediocre after isn't a strategy — it's a cost transfer. Why execution-aware backtesting has to be the default, and what "net of costs" actually has to include.
Code without the data it ran on is a claim you can't check. On pinning datasets to version hashes, and why "it worked last month" is not a reproducible result.
A number has both a date it describes and a date you could have known it. Most datasets only carry the first. Why the second is where look-ahead bias actually hides, and how we store both so a backtest can't quietly use the future.
A risk model that only sees realized volatility misses the drawdown that's still coming. On the limits of variance as a proxy, and what ex-ante risk actually needs to capture.
A split applied a day late, a dividend that never landed. These aren't edge cases — they're the difference between a clean backtest and a silent error that compounds for years. What we check, and why.
Every regime-change model looks brilliant in hindsight and mediocre live. On the lookahead bias hidden in "adaptive" strategies, and how to build regime awareness that only uses information available at the time.
The founding note. Why we started VertiMax, and the definition of rigor we hold ourselves to: no black boxes, no survivorship, and every number traceable to its inputs.
Search enough parameters and you'll find a great backtest by accident. On multiple-testing correction, how the deflated Sharpe ratio prices the search itself, and why we log every parameter we tried instead of only the one that worked.
A −20% drawdown that recovers in a month and one that takes three years are different risks wearing the same label. On time-to-recovery, the underwater curve, and why the peak-to-trough number alone hides the part that hurts.
We ship in public. Each release carries a version number and a record of what changed, so you can pin your research to the platform you actually used.