Free editorial research · Taipei City
Frost Investment Research
Free, plain-language articles about how artificial intelligence is used — and misused — in investing: backtests that look better than they should, and pipelines that quietly leak tomorrow's data into yesterday's model.
No paywall, no signup, nothing is sold. Editorial research only — never investment advice or personalised recommendations.
Why this library exists
A brilliant backtest is a rehearsal, not evidence
Machine learning arrived in investment with a habit the pitch decks rarely mention: it grades its own homework. Train a model on a decade of market history and it can memorise the exam answers instead of the subject. The strategy compounds beautifully right up until it meets real money.
The failure modes have boring names. That is why they survive — they hide inside routine code and ordinary datasets, not in exotic mathematics.
Look-ahead bias
Factors stamped with information that did not exist yet — signals dated Friday that trade on Friday morning.
Train–test contamination
Scaling or feature selection fitted across the whole history, so the test fold shares information with training.
Survivorship bias
Runs on today's index members only — delisted tickers vanish and the graveyard disappears from the sample.
Overfitting to noise
Parameters tuned against one finite past until the graph echoes its accidents rather than its strategy.
What you can read
A small library about how evaluation fails
The article library is the whole offer: free pieces, written to be read end to end, each on one failure mode in the evaluation of AI in investment. No newsletter, no upsell, no login wall — the next essay costs a click.
- Essay 01 · October 2026 Why backtests flatter A backtest is an experiment on the one history we have. Five habits — from lookahead to frictionless fiction — bend that experiment toward the answer you were hoping for. Read →
- Essay 02 · October 2026 Data leakage, catalogued Leakage is any path by which information from outside the training sample — usually the future, often the label — enters model construction. Six named routes, each closable. Read →
- Essay 03 · October 2026 Point-in-time data The dataset you load today is not the dataset a trader saw then. Delistings, restatements, reconstituted indices and recycled tickers all rewrite the past — point-in-time discipline is how the rewrite is undone. Read →
Who this is for — and who should look elsewhere
Read it if you will ever be asked to believe a backtest
Useful to
Engineers and analysts who build or review quantitative models; students meeting financial machine learning for the first time; investors shown a beautiful equity curve who want the vocabulary to interrogate it.
Not offered here
Stock tips, signals, managed money, courses, or personalised advice. Nothing on this site is for sale, no subscriptions or deposits are taken, and inquiries concern the articles — never what you should buy.
If a topic is missing or an explanation is wrong, the inquiry form reaches the editorial desk directly.
Submit an inquiry