Data Machina #261
Substack Carlos
Big names like JPMorgan, Mercedes-Benz, Datadog and Tesco are quietly bolting LLM tricks onto time-series forecasting. Turns out predicting numbers works a lot like predicting words.
Based on reporting by Substack, Carlos — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Time-series forecasting used to be a quiet corner of machine learning, full of ARIMA models and Kalman filters. Not anymore. A fresh batch of research shows big companies racing to graft generative AI techniques onto the business of predicting numbers over time, and the approaches are getting weirder and more inventive by the month.
JPMorgan's contribution, called LETS-C, skips the usual approach of fine-tuning a giant model on your data. Instead it uses a language embedding model to turn time series into vectors, then hands those off to a small CNN-and-MLP classifier. The result beats state-of-the-art classification accuracy while using just 14.5% of the trainable parameters those bigger models need. That's the kind of efficiency win that actually matters when you're running this stuff in production, not just in a paper.
Mercedes-Benz took a different angle with TeVAE, a temporal variational autoencoder built for spotting anomalies in industrial sensor data without needing labeled examples. Tested on real factory data, it caught 65% of anomalies while only misfiring 6% of the time — a solid tradeoff for engineers tired of alert fatigue. Datadog went bigger, training Toto, a general-purpose forecasting model, on one trillion time-series data points, reportedly the largest such dataset ever assembled. Toto now claims state-of-the-art zero-shot results, doing well not just on observability data like server metrics but on unrelated benchmarks too.
Tesco, meanwhile, built a transformer specifically for retail price optimization, using clever tokenization tricks to handle time series at multiple resolutions. It beat the company's existing in-house models in real pricing experiments, which is a rare thing to see stated so plainly in a paper. And then there's ViTime, arguably the strangest idea here: it converts numeric time series into binary images and lets a vision-language model reason over pixel patterns instead of raw numbers. Somehow this beats individually trained supervised models in some tests, proving that sometimes the weird idea works.
Rounding things out is Text2TimeSeries, which throws stock prices, news events, and sentiment data into a T5 transformer to predict short-term price swings. It's a reminder that the line between 'text model' and 'forecasting model' is dissolving fast, and nobody seems bothered by that.
My take — AI-written commentary, not fact-checked reporting
I like this trend more than most AI hype cycles because it's boring in the best way — actual companies solving actual operational problems, not chatbots pretending to be therapists. The pattern worth watching is how fast 'foundation model' thinking is colonizing domains that used to have bespoke statistical tools; forecasting is just the latest one to fall. My only worry is that a trillion-token training run from Datadog quietly becomes the new moat, and open alternatives lag behind in exactly the boring infrastructure areas nobody hypes on Twitter.
Read more about this at: Substack