Taipei · Free editorial · Inquiries only
Granite Engine Research — field notes on factor investing and the craft of feature selection.
We publish free editorial articles on how machine learning reshapes investment research — how factors are constructed, how candidate features are chosen, and where the two disciplines quietly disagree. Nothing is sold here; inquiries are welcome.
Section 01 · The subjects
Where factor research and feature selection meet.
Each piece is built around a concrete problem in quantitative investment research. We write about the methods, the failure modes, and the small choices that change a result.
Factor construction, not factor naming
A factor's name tells you the least. We look at how a signal is sorted, scaled and neutralised — the pipeline that decides what the number actually measures, and why two "quality" factors can disagree.
Feature selection as the real decision
The columns that reach the training set do more work than the algorithm. We compare filter, wrapper and embedded approaches — from mutual information to LASSO and stability selection — and the out-of-sample habits that keep them honest.
Leakage, lookahead and the quiet biases
An over-generous backtest usually hides a feature that peeked at tomorrow. We trace point-in-time data, survivorship and look-ahead bias — the line between a column that forecasts and one that merely recalls.
Section 02 · The craft
How an article is made.
A piece moves through five checkpoints before it is published. The point is not speed; it is that each claim can be traced back to a source or a method.
Frame one question, not a thesis.
We start from a single, answerable question — for example, whether a feature survives out-of-sample once turnover costs are charged. A piece is not published until the question is sharp enough to be wrong.
Separate evidence from opinion.
Academic and practitioner sources are read first, then checked against raw data where we can. Anything we cannot verify is labelled as such, never presented as a confirmed fact.
Write the method before the conclusion.
Before any result is discussed, the method — the universe, the rebalancing, the selection procedure, the metric — is written down in plain language so a reader can reproduce the shape of the work.
Read it as a sceptic.
The draft is re-read with the explicit job of finding the weakest claim. Anything that survives only because of phrasing is rewritten or cut.
Ship it, then keep it honest.
When a later reading or a reader's note changes our view, the piece is corrected and the change is visible, not buried. The archive is a record, not a marketing wall.
Section 03 · Inquiries
Ask about an article, or suggest one.
We do not sell subscriptions, hold client funds, or act as an investment adviser. If a piece left a question open, or you have a subject we should cover, write to us. We read every inquiry and reply by email.
Prefer email? Write to info@granite-engine.digital or call +886 2 2395 6814.