# What's Wrong With The Current Time-Series Deep Learning? ## Provenance - Author: Alexander Chemeris - Primary X status: - X Article ID: `2048011622522781696` - Article created: 2026-04-29 20:00:19 UTC - Article modified: 2026-04-29 20:03:38 UTC - Follow-up X status: - Follow-up created: 2026-05-01 18:37:32 UTC - Snapshot date: 2026-05-16 - Extraction note: article and follow-up text were extracted from public X/FixTweet-style status JSON and X public post text. ## Article TL;DR: Current time-series ML focuses too much on forecasting observations. Foundation time-series models should optimize for maintaining and evolving a useful internal state of the system. ### Forecasting industry bias Outside the field, people hear "time series" and think "finance". Inside the field, people hear "time series" and think "forecasting". Most people don't even think about time-series understanding and reasoning, and most research funding goes into forecasting. But time-series understanding and reasoning have a tremendous economic effect, with applications spanning from network and server management to automating LLM improvements to autonomous nuclear power plants. ### You can't forecast without understanding. Understanding comes before forecasting. In classical models, this understanding is injected by the model designer as an inductive bias through the choice of variables, transformations, assumptions, architecture, loss function, hyperparameters, and training data. ARIMA does not understand the system - but the person choosing ARIMA should. Foundation models change the requirement. We can no longer manually inject the right inductive bias for every task and domain. We want the model to understand the system and the context in which it's operating, and to adapt to the task at hand. We have to teach models to understand before we can teach them forecasting. Understanding should not be an accidental byproduct of forecasting - it should be the training target. ### Multimodality and context importance Providing the right context is a key prerequisite for the model's understanding. The model needs to know the physical meaning of the data it's looking at, and the environment state it's operating in. Luckily, this area has been improving rapidly, with more and more research into providing multimodal context for time series. Referenced arXiv link: ### What's the difference between forecasting and prediction (of the next latent state)? Forecasting is predicting future observations in the target variable space. Whether you can recover the latent state of the system from the observations or not depends on the observation completeness. Prediction of the next latent state of the observed system is more powerful - what regime the system is in, what constraints govern it, what changed, and what futures are plausible. With this, we can derive observations we need or make decisions. This is why I think the field needs to move from observation prediction to state prediction. ### Data scale and distribution Serious enterprise and industrial data spans billions of rows, has hundreds to thousands of dimensions, and is updated in real time. Most widely used benchmarks are still small, static, clean, and low-dimensional compared to real-life time series data. It's been common to complain that models overfit on tiny datasets. Now, you should brace yourself to work with far more data than you can practically use for training. The hard problem is no longer just data scarcity. It is useful-signal scarcity inside oceans of repetitive normal-state data. The most interesting events (failures) are needles in the haystack. Another challenge is maintaining our representation of the system's internal state up to date. The current paradigm of time-series models is not well suited to always-on, real-time operation. PS Thanks to Gemini 3.1 Pro for moral support and GPT-5.5 for valuable feedback on this essay. PPS All of this is written from the perspective of the foundation models. If you work with small task-specific models, you have a different set of challenges. LinkedIn version: ## Follow-Up Post 1/n Wow. I ran a meta-analysis of the @iclr_conf 2026 papers and apparently, 32 out of 57 time-series representation learning papers are from the same EEG/ECG/neuro/physiology space. The rest of time series leans much more toward forecasting. PS I used Codex for batch analysis with some non-exhaustive hand-verification, so some errors might be there, but the result feels correct.