Cadence: Error-Bounded Lossy Compression of Demand Time Series with a Time-Series Foundation Model

Cadence. Error-Bounded Lossy Compression of Demand Time Series with a Time-Series Foundation Model. By Roberto Tacconelli, working independently. Published on the fifth of September. Its code repository is worked out: about forty-five source files, with a readme and a licence.

Cadence compresses numeric time series with loss, but with a hard guarantee: every reconstructed sample stays within a tolerance of the original. It pairs a time-series foundation model of three hundred and thirty million parameters, Google TimesFM-3, with an adaptive arithmetic coder. The obvious hope is that a better forecaster means better compression. One negative result closes that door for lossless coding. There, the bits saved grow only with the logarithm of how much the predictor's error shrinks.

The numbers make it concrete. TimesFM-3 holds a one point five one times accuracy advantage over a linear predictor with thirty-two taps. Under lossless coding, that buys zero point six bits out of twenty point two eight, a median gain of about three hundredths of a percent. For lossless compression, the foundation model adds essentially nothing.

Error-bounded coding escapes this at one point. Once a forecast lands inside the tolerance band, the residual index is zero, and that sample costs almost nothing to store. A more accurate forecaster lands inside the band more often, and there accuracy pays in a way the logarithm does not allow.

The system makes three contributions around that idea. First, an adaptive range coder with context-modelled binarization, which beats xz and zstd on real indices by nine point seven percent, in fifteen cases of fifteen, reversing a finding obtained with a general-purpose back end. Second, a determinism result. Predictions are not bit-identical across batch sizes, and no PyTorch configuration repairs that, so group size and execution device have to be written into the container format. Third, domain localization, tested on corpora that postdate any plausible training cutoff.

On forty-nine EIA-930 balancing-authority demand series, Cadence gains thirteen point three percent over the best of six classical predictors. On fifty MTA ridership series it gains twenty-eight point three percent. Across two hundred and ninety-seven series and tolerance pairs, the median gain is twenty-one point four percent, and Cadence wins all of them.

Against downsampling, which is what time-series databases deploy for retention, its guaranteed worst-case error is twenty-eight to fifty-six times tighter at equal size. End to end, once the context bootstrap is paid for, gains run from six point eight percent at six months of hourly data to fifteen point one percent asymptotically.

The authors also tried to falsify the domain claim, on SDRBench. Theory predicted failure there, and failure is what came: a median of minus zero point eight percent, with none of twenty-seven pairs gaining. Three further negative results and eight retracted claims are reported in full.

The paper's argument runs through its negative results as much as its gains. A foundation model does not help lossless coding, and cannot, because of how bits scale with accuracy. It does help when a tolerance turns good forecasts into near-free samples, and it stops helping outside the domain it knows, just as predicted.