Time-Series JEPA¶
JEPA, put on a clock: learning the structure of time series by predicting the future in latent space.
Most of machine learning's progress on images came from a simple idea: hide part of the data and learn to predict it. JEPA — the Joint-Embedding Predictive Architecture — does this in a particularly clean way, predicting the meaning (the embedding) of a hidden region rather than its raw pixels. Time-Series JEPA is that same idea pointed at time: hide the future, and learn to predict it from the past — again in latent space.
This series is a self-contained tutorial on Time-Series JEPA as a method in its own right. It is the gentlest entry point in this corpus: if you know roughly what self-supervised learning is, you have enough to start. You do not need any of the action-operator or "world model" material to follow it — that comes later, and Time-Series JEPA is what it will be built on.
What you will get¶
A working mental model of how to learn from unlabeled time series, told as one continuous story:
- From I-JEPA to Time-Series JEPA — how "predict a masked patch of an image" becomes "predict the next stretch of a time series," and why doing it in latent space (not raw signal) is the right call.
- Multimodal temporal channels — the real-world complication: several signals at once (a wearable, a phone), sampled at different rates, often missing. Handling this is the research problem, and it is where Time-Series JEPA earns its keep.
- The limits of pure Time-Series JEPA — one honest blind spot that pure prediction cannot fix, which is exactly the door into the rest of the corpus.
Throughout, we follow one running example — a person's day-to-day behavioral rhythm — so each idea lands on something concrete you can picture, not just notation.
Before you start: what to read, and what you can skip¶
You need only a passing familiarity with self-supervised learning and the original I-JEPA (image JEPA). We recap the essentials in Part 1 in fresh words, so a quick memory is enough; for a fuller treatment see the JEPA encoder chapter in the Generative JEPA series.
You can skip, for now, the Action Operators foundation and the Operator World Models series. Time-Series JEPA stands on its own. Those become relevant only once we hit the blind spot in Part 3 — at which point this series hands you forward to them.
Reading order¶
| Part | Topic | What you get |
|---|---|---|
| 1 — From I-JEPA to Time-Series JEPA | the temporal reinterpretation; the four pieces; the forward pass | how masked-patch prediction becomes next-step prediction, with a worked example |
| 2 — Multimodal temporal channels (coming next) | async, irregular, missing streams; per-modality encoders; masking curricula | how Time-Series JEPA handles many real signals at once |
| 3 — The limits of pure Time-Series JEPA (coming next) | the "what acted?" blind spot; the contaminated surprise signal | why pure prediction is not enough — and where to go next |
New to a symbol? The notation reference defines every one.
Where this leads¶
Time-Series JEPA gives you a model that predicts what comes next. By Part 3 you will see that it cannot, on its own, say what acted to bring that next state about — and for problems like deciding which intervention helps a patient, that gap is everything. Closing it is the job of the Action Operators foundation and the Operator World Models series, for which this one is the natural runway.
Related work, and what is new here¶
Applying JEPA to time series is an idea several groups have arrived at independently — which is a good sign it is a natural step, not a niche trick. For the published lines, see S-JEPA (Signal-JEPA, an early JEPA applied to EEG signals, arXiv:2403.11772), TS-JEPA / Joint Embeddings Go Temporal (arXiv:2509.25449), and a Time-Series JEPA applied to predictive remote control (arXiv:2406.04853). For the multimodal setting of Part 2, see Giving Sensors a Voice: Multimodal JEPA for Semantic Time-Series Embeddings (OpenReview), which introduces CHARM, a channel-aware multimodal model.
A naming note: the acronym T-JEPA in the literature refers to unrelated work on tabular (arXiv:2410.05016) and trajectory (arXiv:2406.12913) data — which is why this series uses the descriptive name Time-Series JEPA throughout.
This series is an accessible, example-driven exposition of that shared foundation. Its distinctive contribution is not the temporal predictor itself, but what we build on it: conditioning the predictor with an action operator — the synthesis developed in the Action Operators foundation and the Operator World Models series, which goes beyond the prediction-only models above.
Start with Part 1 — From I-JEPA to Time-Series JEPA.