Cadence: Error-Bounded Lossy Compression of Demand Time Series with a Time-Series Foundation Model
Cadence gets its edge when forecasts fall inside the allowed error band, making those samples nearly free to encode.
The paper pairs Google TimesFM-3 with an adaptive arithmetic coder while guaranteeing per-sample error stays within `τ`. It reports 13.3% gains over six classical predictors on 2026 EIA-930 demand series and 28.3% on 2026 MTA ridership series. The author also says foundation models do not help much for lossless coding, because better prediction accuracy converts into only small bit savings. A test on SDRBench fails as predicted, with a -0.8% median result and no gains across 27 pairs. HF Daily Papers' note
The paper pairs Google TimesFM-3 with an adaptive arithmetic coder while guaranteeing per-sample error stays within `τ`. It reports 13.3% gains over six classical predictors on 2026 EIA-930 demand series and 28.3% on 2026 MTA ridership series. The author also says foundation models do not help much for lossless coding, because better prediction accuracy converts into only small bit savings. A test on SDRBench fails as predicted, with a -0.8% median result and no gains across 27 pairs. HF Daily Papers' note
score 4