|
|
|
24 July 2026
Online version
AI Without Discipline Is Just Curve-Fitting
|
Every other week, a new paper shows LLMs can "predict" inflation, markets, or both. The same two pitfalls usually go unmentioned.
The first one is the word "sentiment". It remains the default label for anything derived from text, often for lack of a better word (and, we suspect, of a clearer purpose). We wrote about this back in 2020 (link): build a "sentiment" indicator without removing the news about the target variable itself, and what you get is essentially the target variable lagged by a day or so. The correlation is spurious. Whose sentiment, then? Know your data.
The second one is more recent: look-ahead bias, LLM edition. A general-purpose model released in 2026 has read the whole 2022-2025 inflation story. It will "forecast" that period very well, because it remembers it. Any backtest where the LLM training sample overlaps the evaluation window is contaminated by construction, and with foundation models the overlap is more or less guaranteed. The truth is, that problem does not go away with scale. It gets worse.
Our answers to these challenges have not changed: narrowly targeted models, a use case defined before writing any code, and a track record that has been live and truly out-of-sample since 2020.
In this issue: a puzzling chart on core inflation pressures; Themes, our new portal feature; and the July Global NIPI report.
Enjoy the reading!
|
|
|
|
|
|
|
|
Chips, China, and Everyone Else's CPI
China's Core NIPI has surged back above 120, now well ahead of the US signal, and one theme explains most of it: electronics, where AI demand for memory chips is crowding out consumer supply and lifting prices from DDR5 kits to flagship smartphones. This is worth more attention than a domestic Chinese story would deserve. China sits at the head of global supply chains, and nothing about this shock is China-specific — the demand is global, and so are the products. There is no reason for it to stay local. Our Themes heatmaps flagged the narrative weeks ago; the transmission is the part to watch now.
|
|
|
|
|
|
Introducing Themes and Themes Heatmap
Themes is the newest addition to the AMS portal: daily heatmaps showing which topics are driving inflation news across ten economies, in which direction, and how that has shifted over the past month. The launch post walks through two live examples, from memory chips lifting electronics prices in China to tariff pass-through in the US.
|
|
|
|
|
|
July-26 Global monthly NIPI Report
The July Global NIPI report is out: a monthly, news-based read on inflationary pressures across 23 economies, published ahead of official statistics. The report is free, the data can be downloaded, and you can register here to receive it by email.
|
|
|
|
|
|
|
|
|
|
|
|
Earlier newsletters
Blog Post: Themes Heatmaps: Inflation Narratives at a Glance, 21/7/2026. Available here. Global NIPI Monthly Report: July 2026. Available here. Newsletter: AMS H1 Recap: NIPI forecasting results, optimization, and NewsBot updates, 4/6/2026. Available here. Blog Post: Tweaking the NIPI parameters to improve inflation forecast accuracy, 2/6/2026. Available here. NewsBot Feature Update: NewsBot Feature Updates: Themes, Spotlights, and Portal, 19/5/2026, here (restricted).
|
|
|
|
|
|
|
|
|
|
|
