Articles
Can Large Language Models Forecast Time Series?
"Can't today's AI already do time series forecasting?" The push to forecast time series with LLMs began in 2023 and is still unsettled in 2026. Can they, or can't they? A time series researcher separates what the papers actually showed from personal opinion.
→What Is a 1% Gain in Forecast Accuracy Actually Worth?
"We improved MAE by 5%." And how much money is that worth? The link between forecast accuracy and value has been studied since 1951. Minimizing an error metric and optimizing a decision are not the same problem.
→Forecasting JEPX Prices with Time Series Foundation Models: The Baseline to Beat Was Yesterday (Part 1)
"Isn't a classical statistical model enough for electricity price forecasting?" European studies suggest so. A year of zero-shot evaluation across ten JEPX spot price series shows the baseline to beat is simply copying yesterday, and only the time series foundation models beat it. Importing the standard European benchmark design leads to the wrong conclusion.
→The Evolution of Irregular Time Series Models: From GRU-D and Neural ODE to mTAN and Beyond
"Can't you just fill in the gaps before feeding the model?" Some might think so. From GRU-D and Neural ODE to mTAN and ISTS-PLM, research on irregularly sampled time series has shifted from how to impute missing values to how to represent irregular time inside the model. This article traces that shift.
→Why Ordinary Models Cannot Forecast Irregularly Sampled Data
"If the sampling is irregular, just resample to a regular grid before feeding the model." That is half right: ordinary models cannot take irregular data as is, so without resampling the accuracy does not come out. But resampling measures accuracy on data from which the timing of each observation has been thrown away, and that distortion never shows up in the accuracy number. This article checks both claims with synthetic data and PhysioNet.
→Forecasting Power Demand with Prophet and Time Series Foundation Models: What Mattered Was Temperature, Not the Model
"Isn't Prophet enough for demand forecasting?" Some might think so. On nine years of daily peak demand in TEPCO's service area, Prophet did not significantly beat simply copying the value from seven days earlier. What mattered was not the choice of model but how temperature was fed in.
→The First Step Toward Time Series Forecasting: OLTP and OLAP
Running aggregation queries against the production database makes them compete with ordinary requests for CPU and disk. How do you pick an analytics database, and how do you get data into it? This post lays out four sync methods and four stores, then benchmarks the same query over 11.68M rows across five setups.
→Slam Dunk Was Read More on Announcement Day Than on Release Day: Four Shapes of Booms in Wikipedia Pageviews
On the day THE FIRST SLAM DUNK opened, the Japanese Wikipedia article on Slam Dunk was read 7.9 times as often as usual. On the day the film was announced, it was 21.8 times. Measuring the shape of attention across 36 items, from obituaries and anime to buzzwords and Heisei-era revivals, yields four types. The arrival of a peak cannot be read from the views before it, but the decay after a peak can be roughly read from its first seven days.
→Study & Survey
Latest research surveys and study session materials.