close
One open time series model,
OTIS is an open, general-purpose time series model: pre-trained once, it can be used frozen on data it has never seen, from ECGs and EEGs to power meters and
light curves, or fine-tuned in a blink.
It is tiny. 7 million parameters in 7 megabytes, exported to ONNX, quantised to int8 and run with
onnxruntime-web in WebAssembly. Everything is run locally on your CPU. Nothing leaves your browser.
Comparable models weigh a hundred times more. This page lets you feel the difference: same data, same CPU, different runtime.
Turgut, Müller, Menten & Rueckert. OTIS: Learning High-Quality Time Series Features With Tiny Encoders. TMLR 2026.
Explore
The time series model encodes every recording without ever being told what the recording is. The map
places recordings with similar encodings close together; the colours are the true classes, which the time series model never
saw. Hover a dot to see the recording and click it for the details. Each panel shows three numbers: the time the model took
for all recordings on your CPU, its throughput in recordings per second, and the accuracy, how often a test recording’s
nearest training recordings share its class, so higher is better and guessing would score about one over the number of
classes. Click an empty panel to run the time series model on your CPU.
Setup: 200 training and 100 test recordings per dataset, balanced over the classes, each encoded once by the frozen time
series model with no fine-tuning. Accuracy is the balanced accuracy of a weighted nearest-neighbour vote over the training
recordings, so every class weighs the same; the map is a t-SNE of the same encodings and the labels only colour it. The
training encodings were computed with the same graphs. Your data runs the same protocol and nothing is
uploaded. Data: PTB-XL, Sleep-EDF, BasicMotions, StarLightCurves and SmallKitchenAppliances; the protocol and the full results
are in the paper. The first click on a panel fetches the time series model. The test recordings are encoded on your CPU.
Impute
Pick a recording (a sine, chirp, square wave or random walk with the frequency and noise of your choice, or
draw or paste your own), then hide part of it: at random, or by dragging across the chart. The hidden part is shown in grey.
Each time series model gets the visible part only and has to predict the hidden steps; the chart lays its prediction over the
truth. Click Run on a row: the time series model runs on your CPU and the row shows how long it took, then MSE and MAE, the
distance from the truth on the hidden steps, lower is better and MSE weighs large misses more, and NCC, how well the shape
matches, from 1 (perfect) through 0 (no relation) to -1 (the inverted shape).
Setup: a 432-step recording, z-scored, with the hidden steps masked. Every time series model sees the same visible steps,
never the hidden ones, each on its own patch grid, so the span is rounded to it. Each is frozen and runs its own
masked-modelling path; only its few task parameters were fitted, all three on the same signals: sines, chirps or squares drawn
afresh between 20 and 120 Hz at every step, so none is tuned to the rate you pick. OTIS fitted the fewest, 96 parameters, and left
its mask token and head frozen; UniTS fitted 1,536 and MOMENT 9,224, both including theirs. For a random walk, a drawing or
pasted numbers, all three use the sine's. MSE, MAE and NCC
are computed on the hidden steps only. Changing the recording or the hidden
span clears the rows. The first click on a row fetches the time series
model. The prediction is computed on your CPU.
Forecast
Pick a recording and one of its windows; on a recording with several variates, the dropdown picks the one
plotted. Each time series model sees the past of the recording and predicts the steps that follow, the shaded part; the truth
is drawn for comparison. Click Run on a row: the time series model runs on your CPU and the row shows how long the window
took, then MSE and MAE, the distance from the truth on the predicted steps, lower is better and MSE weighs large misses more,
and NCC, how well the shape matches, from 1 (perfect) through 0 (no relation) to -1 (the inverted shape).
Setup: four standard benchmarks (Weather, Illness, Electricity, Traffic), split and z-scored as in the paper, six test
windows each. Every time series model sees the same 336 steps and predicts the next 96 (48 for Illness), all variates at once;
the time is for all variates, the scores for the variate plotted, in the z-scored units of the dataset. A window’s
results are kept while you move between windows and recordings. The first click on a row fetches the time series model. The
forecast is computed on your CPU.