Dynamic Topic Models for Market Narratives

A standard topic model, run once on a year of financial news, will happily tell you that "inflation," "supply chain," and "rate hikes" were common themes over the whole period — but it treats the entire year as one static bag of documents, blind to the fact that "supply chain" dominated the conversation in one quarter and had all but disappeared by the next. Markets don't just react to what is being discussed, they react to how the discussion is shifting — a narrative gaining share of voice week over week is informative in a way a single year-end topic summary can never capture.

A dynamic topic model extends ordinary topic modeling (see Topic Modeling with LDA) by fitting the topic structure separately, but smoothly, at each point in time — typically week by week or month by month — so that both the words that define a topic and the share of documents devoted to it are allowed to drift, while remaining linked to their earlier values rather than being estimated from scratch each period.

What changes over time, and how it's used

Two things a dynamic topic model tracks that a static one cannot:

• Topic prevalence — the fraction of documents each period that are about a given topic. A rising share for "AI capital expenditure" across successive months is a directly measurable proxy for how much the market's attention has rotated toward that theme.
• Topic content drift — the specific words most associated with a topic can shift even while the topic's broad label stays the same; "inflation" discussed in terms of energy prices one quarter and in terms of wage growth the next is the same topic label carrying a materially different narrative.

In practice a desk tracks prevalence time series for a fixed set of topics and treats a topic's week-over-week change in share — not just its level — as the feature, since a narrative accelerating in attention behaves differently from one at a stable, already-priced-in level.

Worked example

A dynamic topic model fit monthly over two years of macro-relevant news identifies a persistent "supply chain disruption" topic. Its estimated prevalence sits around 4% for most of the sample, then rises to 9%, then 15%, then 22%, before declining back toward 5% over the following four months. A desk treating month-over-month prevalence change as a feature would flag the first two months of the rise as the most informative window, since that's where the narrative was accelerating fastest — by the time prevalence peaked at 22%, the theme was common knowledge and largely priced in, and the decline months carried little fresh information.

What this means in practice

Dynamic topic models give a desk a systematic way to answer "what is the market newly focused on" rather than relying on a human reading headlines and guessing, and the prevalence time series can be backtested like any other feature against subsequent returns. The main practical cost is that re-fitting the model each period requires care: topics need to be matched across periods (topic 7 this month must be linked to the equivalent topic last month) or the prevalence series becomes meaningless, and a topic definition that drifts too much period to period stops being a stable feature at all.