topics covered by the desk
The reading list: threads, not titles
No paywall and no product behind it: the desk publishes about AI in investment — portfolio diversification and correlation analysis — in four slowly growing threads. Each thread below lists what the reading inside it picks apart.
How clustering quietly rebuilds diversification
The oldest number in portfolio theory is “how many different things do I hold?” Machine learning reframes it as “how many different behaviours do I hold?” Articles in this thread walk that reframing without worship.
- how clustering algorithms regroup an index universe by return behaviour instead of sector labels
- hierarchical trees (dendrograms) read from the leaves: which holdings fatten the tail while pretending to be different
- concentration measures that notice ten positions behaving as four
- equal-weight vs. risk-parity vs. cluster-weighted mixes, redrawn on public data each time
Correlation as a moving estimate
A correlation matrix is a snapshot of a moving object; read as a still, it flatters. This thread charts the motion.
- rolling-window versus exponentially weighted estimators, and how window length bends the answer
- regime shifts: the quiet jumps that re-price diversification long before headlines do
- spillover: how a session in one market carries into the books of another that slept through it
- heatmaps audited honestly — where colour compresses exactly the information you needed
What the code carries unspoken
Estimators carry their priors in their pockets. This thread empties the pockets onto the table.
- shrinkage priors and factor structure weighed against plain sample covariance, in arithmetic you can follow
- walk-forward testing explained as a habit, not a ritual word
- what a neural network buys in covariance estimation — and what it charges in opacity
- stationarity: the lie every estimator is encouraged to tell, and when to stop forgiving it
Reading AI claims about markets
A media-literacy thread: how to interrogate a claim that “AI drives performance” — in a fund pitch, a paper abstract or a panel quote — in under ten minutes.
- separating what the model was, what the data was, and what the release letter wanted
- backtest hygiene: sample windows, survivorship and the unlabelled winners
- three questions that unmask most correlation-estimation marketing
- where published methods contradict their own marketing copy, we say precisely where
Reader questions, answered without hedging
Is any of this investment advice?
No. Everything here is general commentary on published methods and public data. Nothing is personalised, nothing recommends a security, and no question you send will be answered with a position call. For decisions about your own money, a licensed adviser acting for you is the right counterparty.
Does the desk sell anything?
Nothing at all: no subscriptions, no paid reports, no courses, no services, no premium tier. The articles are free, the inquiries are free, and the email address costs a stamp of nothing to write to.
Where do the data in articles come from?
Public, checkable sources: exchange disclosures, index fact sheets, published papers and freely available price histories. Every article names its data and its window, and charts are redrawn from raw numbers rather than reused screenshots.
Can I request a topic?
That is exactly what the inquiry form is for. Tell the desk what you want unpacked — a clustering method, a correlation puzzler, a suspect chart — and it joins the queue, in the order topics are finished, not in the order they would sell… because nothing sells.
Request a topic for the queue
One form, read by one desk, answered by the same person who writes the pieces.