{"id":1916,"date":"2026-04-26T09:14:35","date_gmt":"2026-04-26T09:14:35","guid":{"rendered":"https:\/\/inphronesys.com\/?p=1916"},"modified":"2026-04-26T09:42:14","modified_gmt":"2026-04-26T09:42:14","slug":"why-var-beat-googles-timesfm-and-how-to-build-one-in-r","status":"publish","type":"post","link":"https:\/\/inphronesys.com\/?p=1916","title":{"rendered":"Why VAR Beat Google’s TimesFM \u2014 and How to Build One in R"},"content":{"rendered":"

In December 2025, researchers at UC San Diego published a peer-reviewed receipt on something the forecasting community has been tiptoeing around for two years: Google’s shiny new time-series foundation model, TimesFM<\/strong>, got beaten by a model older than the first iPhone. By a lot. The winner \u2014 vector autoregression (VAR)<\/strong> \u2014 has been kicking around since 1980, when Christopher Sims (the eventual 2011 economics Nobel laureate) wrote the paper that made it famous. The score was not close.<\/p>\n

At the La Jolla emergency department, forecasting how many patients would be stuck „boarding“ (admitted but not yet in a ward bed), VAR scored an RMSE of 8.76<\/strong> at the one-day-ahead horizon where it was actually deployed. Google’s TimesFM, at the same T+1 horizon, scored 14.54<\/strong> \u2014 66% worse than VAR<\/strong>. The 2-week moving average baseline sat at 14.85<\/strong>, which means VAR cut error by 41% versus the naive benchmark<\/strong> \u2014 while TimesFM failed to beat that benchmark at any<\/em> horizon, getting worse as the forecast stretched (RMSE 16.88 at T+4). Out of six candidate models, VAR was the only one the hospital put into live production, quietly emailing forecasts to Mission Control every morning at 7:10.[^1]<\/p>\n

If you run a supply chain, that result should stop you mid-coffee. Because the reason VAR won in a hospital is exactly the reason it tends to win in demand planning too \u2014 and almost nobody in SCM is using it. This post walks through what VAR actually is (without the matrix algebra), why supply chain data is practically designed for it, and how to build one in R in about thirty lines of code.<\/p>\n

What VAR Actually Is<\/h2>\n

Here’s the one-sentence version: VAR is ARIMA with friends.<\/strong><\/p>\n

An ARIMA model stares intensely at one time series \u2014 weekly demand, for example \u2014 and tries to predict its next value from its own past. It’s a monk. Deeply introspective. Ignores everything else in the room.<\/p>\n

A VAR model, by contrast, watches a small group<\/strong> of related time series at the same time and lets them gossip<\/em>. Each series predicts its next value using its own past AND the past of every other series in the group. Weekly demand depends on last week’s demand, and<\/em> on last week’s price, and<\/em> on last week’s promotion intensity. Price depends on its own past and on demand. Promotions depend on both. Everyone tells on everyone else.<\/p>\n

That’s the whole idea. No matrices, no eigenvalues, no doctoral-defence nightmares. Just: if you’ve got three or four series that you suspect talk to each other, VAR lets them actually talk.<\/p>\n

The quiet genius of Sims’s 1980 paper „Macroeconomics and Reality“<\/em>[^2] was philosophical. He argued that economists were imposing too much theoretical structure on their models \u2014 deciding in advance that „interest rates cause inflation“ and baking that into the equations. His counter-proposal: let the data decide who causes whom<\/strong>. Treat every variable symmetrically. Regress each one on lagged values of itself and<\/em> all the others. Read the tea leaves afterwards. Central banks \u2014 the Fed, the ECB, the Bank of England \u2014 still use VARs and their descendants as their workhorse short-term forecasting tools. Not because they’re glamorous. Because they work.<\/p>\n

Why Supply Chain Data Is Natively Multivariate<\/h2>\n

Here’s the awkward truth about classical demand forecasting: we pretend demand exists in a vacuum. You look at a weekly sales series, fit an ETS or ARIMA, and call it a forecast. But nobody in purchasing actually believes the sales number moves on its own. What moves it?<\/p>\n