(Original → Compressed)
Error Bound
Error (MAE)
Error (RMSE)
Percentage Error (MAPE)
Squared Error (NRMSE)
For logical-level compression methods, the percent sweep is interpreted by category: maximum per-point deviation for most compression methods, RMSE for Bottom-up and Sliding-window, and area error for Visvalingam–Whyatt. DFT uses retained coefficients as a percentage of series length; LTTB uses target points as a percentage of series length. Both size sweeps run left-to-right from lower compression to higher compression, so the retained percentage decreases.
Camel vs BitPackedBUFF swept over decimal precision 1–4 (Camel's maximum). Their compression ratio is low next to the other physical compression methods, so they're easy to miss in the shared chart above. If your series values already have fewer than 4 decimal places, Camel reconstructs them exactly (lossless, RMSE ≈ 0) at that precision and above — so it collapses to a single point in the shared chart above and flattens into a constant line here.
Compose a three-stage pipeline. The logical compression method approximates the series as segments. The indices compression method stores the segment positions exactly. The coefficient compression method compresses the numeric segment values used to reconstruct the curve. The table compares the combined pipeline against the logical and coefficient compression methods applied to the raw series on their own.
Optimize searches for compression pipelines composed of a logical compression method, an indices compression method, and a coefficient compression method for the loaded time series. Find out which combination of methods and error bounds best balance compression ratio and decompression error. The optimization process evaluates many combinations of methods and error bounds, ranks them by a reward function, and presents the top results in a table.
Work in progress. Optimize is experimental and may take a few minutes depending on the series length. Results appear automatically as they become available.