Batch time-series feature extraction: 33 features at 3.18 ms median per 1,000 series x 500 steps, with zero defensive memory copies.
Kymora is a minimalistic, dependency-light time-series feature extraction library designed to make extracting statistical, temporal, and spectral features across large datasets fast and memory-efficient. It combines a zero-copy Rust engine with a clean, Scikit-Learn-compatible Python interface—ideal for machine learning pipelines, quantitative finance, real-time sensor telemetry, and high-throughput research.
PyPI • Features • Installation • Quickstart • Benchmarks • Streaming & Real-Time Telemetry • Scikit-Learn Integration • Documentation • Website
Traditional Python time-series feature libraries (tsfresh, TSFEL, catch22) force a painful trade-off: wait minutes to hours for feature extraction, or risk Out-Of-Memory (OOM) crashes from defensive copies. Kymora eliminates that trade-off.
-
Throughput (measured): Best run 436,719 series/second (median 314,450) on an i7-13620H laptop, 10 cores / 16 threads — about 262×
catch22and 6,570×tsfreshraw time on the same machine (per-feature: 393× and 279× — see Benchmarks andCLAIMS.md). - Zero-Copy Ingestion: Directly borrows contiguous NumPy buffer pointers via PyO3. No data duplication, no DataFrame melting, and zero intermediate memory ballooning.
-
33 Curated, High-Signal Features: Avoids the curse of dimensionality. Features are mathematically non-redundant (
$|r| < 0.70$ for 83.3% of pairs), spanning distribution moments, quantiles, crossings, spectral power, and permutation entropy. - Full Multi-Core Scaling (GIL-Free): Releases Python's Global Interpreter Lock (GIL) across the entire computation region, saturating all CPU cores with Rayon's work-stealing scheduler.
-
Real-Time Streaming Ready: Maintain rolling windows with incremental state updates using the built-in
StreamingExtractor— amortized$O(1)$ per-sample ingestion plus a fast 12-feature tier with no sorting and no FFT.
- Zero-Copy Hybrid Architecture: PyO3 bindings pass 2D NumPy pointer references directly into native-Rust multi-core loops without copying a single byte.
-
Batch-First Parallelism: Processes
$N$ series in parallel across hardware threads instead of running serial Python loops. - Dual API Support: Extract raw 2D NumPy matrices for maximum speed, or labeled Pandas/Polars DataFrames for immediate exploratory analysis.
-
Scikit-Learn Compatible: Drop
KymoraTransformerinto anysklearn.pipeline.Pipelineor cross-validation grid search. - Realfft & Histogram Quantiles: Preallocated per-worker FFT scratch and histogram multi-select quantiles ensure predictable sub-millisecond execution.
- Streaming & Sliding Windows: Extract rolling features over continuous data streams without reallocating buffers.
Precompiled binary wheels are available on PyPI (kymora) for Linux (x86_64, aarch64), macOS (Apple Silicon arm64, Intel x86_64), and Windows (x64). No Rust compiler required!
pip install kymoragit clone https://github.com/Aamodx6/Kymora.git
cd Kymora
pip install maturin
maturin develop --releaseExtract 33 features from 100,000 series in under a second:
import numpy as np
import kymora as km
# 1,000 series of 500 time-steps (float64)
X = np.random.randn(1000, 500)
# Extract 33 features (zero-copy, multi-threaded)
features = km.extract_features(X)
print("Output shape:", features.shape) # (1000, 33)
print("Feature names:", km.feature_names()[:5])
# ['mean', 'std', 'var', 'min', 'max', ...]# Returns a labeled pandas DataFrame with clean column headers
df = km.extract_features_df(X)
print(df.head())# Sequences of varying lengths are supported natively
arr1 = np.random.randn(300)
arr2 = np.random.randn(500)
arr3 = np.random.randn(120)
features = km.extract_features([arr1, arr2, arr3])
print(features.shape) # (3, 33)# Extract rolling window features from a 1D continuous sensor stream
signal = np.random.randn(100_000)
windowed_features = km.sliding_features(signal, window=256, stride=64)Integrate directly into standard classification, regression, or clustering pipelines
(pip install "kymora[sklearn]" for the optional dependency):
import numpy as np
from kymora.sklearn import KymoraTransformer
from sklearn.pipeline import Pipeline
from sklearn.ensemble import RandomForestClassifier
from sklearn.preprocessing import StandardScaler
# Assemble end-to-end reproducible pipeline
pipeline = Pipeline([
("features", KymoraTransformer()), # (n_series, length) -> (n_series, 33)
("scaler", StandardScaler()),
("classifier", RandomForestClassifier(n_estimators=100))
])
# Synthetic training data (n_samples, time_steps) — replace with your series
rng = np.random.default_rng(0)
X_train = rng.standard_normal((40, 128))
y_train = (X_train.sum(axis=1) > 0).astype(int)
X_test = rng.standard_normal((10, 128))
# Fit on raw time-series training data (n_samples, time_steps)
pipeline.fit(X_train, y_train)
y_pred = pipeline.predict(X_test)Maintain running statistical features in real-time embedded systems or trading loops without recomputing from scratch:
import numpy as np
from kymora import StreamingExtractor
# Initialize streaming extractor with window capacity
stream = StreamingExtractor(window_size=500)
incoming_data_feed = iter(np.random.default_rng(1).standard_normal(600))
for tick in incoming_data_feed:
stream.push(tick)
# 1. Fast 12-feature tier (O(1) accumulator reads — no scan, no sorting, no FFT):
# Returns 12 features: mean, std, var, skew, kurt, abs_energy, rms,
# mean_abs_change, mean_change, cid_ce, zero_crossings, trend_slope.
# Accuracy matches batch within rtol 1e-9 (see docs/streaming.md for the
# large-offset skew/kurt conditioning note).
fast_features = stream.compute(kind="fast")
# 2. Complete 33-feature set evaluated over the current rolling window:
all_features = stream.compute(kind="all")Choose the performance-to-breadth profile that fits your pipeline:
minimal(10 features): Centered moments, extrema, energy, zero crossings. Zero sorting and zero FFT overhead (~1.20 ms per 1,000 series; 833,333 series/s, re-measured 2026-10-05).core33(33 features - Default): Frozen authoritative v1.0 set spanning all temporal, quantile, and spectral domains (~2.88 ms median per 1,000 series on i7-13620H / 16 threads, re-measured 2026-10-05; the 2026-10-04 F1 pool measured 3.18 ms / 314,450 series/s — run-to-run spread, both artifacts committed; 0.0964 µs per series-feature).extended(143 features): Adds distribution statistics, crossings, nonlinear stats, PACF (Levinson-Durbin), full linear regression trend, and spectral aggregations.full(543 features): Complete high-coverage bank including all 400 FFT coefficient parameters extracted directly from the precomputed spectrum with zero redundant transforms.
import kymora
# List available profiles and feature counts
print(kymora.list_profiles())
# {'minimal': 10, 'core33': 33, 'extended': 143, 'full': 543}
# Inspect individual features and their computational prerequisites
print(kymora.describe_feature("autocorr_lag_1"))All numbers below are pooled medians from interleaved rounds with 95% bootstrap CIs, subprocess-isolated, GC disabled — from one laptop (i7-13620H, 10 cores (6P+4E) / 16 threads, Windows 11, Performance plan, AC online). Exploratory single-machine numbers, not fleet evidence. Claim→artifact map: CLAIMS.md. Artifacts: benchmarks/results/F1_REPORT.md (raw-time + competitor rounds, 2026-10-04), benchmarks/results/EQUAL_FEATURE_REPORT.md + 2026-10-05_equal_feature/ (equal-feature, 2026-10-05), benchmarks/results/2026-10-05_remeasure/ (profiles / scaling / memory, 2026-10-05):
Both sides compute exactly the same features: each library is restricted to the subset whose definitions match kymora's, and every feature pair is numerically verified before timing (parity gate: max rel err ≤ 1e-9; artifact 2026-10-05_equal_feature/parity.json). 16 threads:
| Library | Equal features | Kymora med (ms) | Library med (ms) | Ratio (95% CI) |
|---|---|---|---|---|
| numba baseline | 33 | 4.72 | 16.05 | 3.4× slower [3.1, 3.9] |
| numpy baseline | 33 | 4.68 | 285.50 | 61.0× slower [60.0, 62.0] |
| TSFEL | 13 | 3.92 | 2,213.69 | 564.6× slower [509.7, 781.4] |
| tsfresh | 13 | 4.94 | 8,359.45 | 1,693.8× slower [1,645.5, 1,900.1] |
Kymora is fastest on all 16 equal-feature rows measured across 4 shapes (ratios 3.4×–1,693.8×; full tables: benchmarks/results/EQUAL_FEATURE_REPORT.md). catch22 has no row here by design: none of its 22 features is definition-identical to a core33 feature (mapping and near-miss analysis: docs/benchmarks/feature_mapping.md).
| Profile | Features | Latency (1k) | Per-Series | Per-Feature Cost | Throughput |
|---|---|---|---|---|---|
minimal |
10 | 1.20 ms | 1.20 µs | 0.120 µs | 833,333 series/s |
core33 |
33 | 2.88 ms | 2.88 µs | 0.0873 µs | 346,963 series/s |
extended |
143 | 10.10 ms | 10.10 µs | 0.0706 µs | 99,028 series/s |
full |
543 | 11.10 ms | 11.10 µs | 0.0204 µs | 90,052 series/s |
Artifact: benchmarks/results/2026-10-05_remeasure/profiles.jsonl (medians of subprocess-isolated runs per profile).
| Worker Threads | Latency | Per-Series Cost | Speedup vs 1 Thread | Scaling Efficiency |
|---|---|---|---|---|
| 1 Thread | 12.47 ms | 12.47 µs | 1.00× | 100.0% |
| 2 Threads | 7.01 ms | 7.01 µs | 1.78× | 89.0% |
| 4 Threads | 4.11 ms | 4.11 µs | 3.03× | 75.7% |
| 8 Threads | 3.03 ms | 3.03 µs | 4.12× | 51.5% |
| 16 Threads | 2.93 ms | 2.93 µs | 4.25× | 26.6% |
| 32 Threads | 3.51 ms | 3.51 µs | 3.56× | 11.1% |
Efficiency drops past ~4–8 threads on this hybrid P+E-core laptop (oversubscription past 10 physical cores); treat the shape of this table as hardware-bound. Artifact: benchmarks/results/2026-10-05_remeasure/scaling.jsonl.
Each library times its full default catalog end to end, including the input reshaping it requires (tsfresh's long-format DataFrame, catch22/TSFEL's per-series loop):
| Library | Features | Runtime (1k × 500) | Series / sec | Speedup (raw time) | Per-feature | Speedup (per-feature) |
|---|---|---|---|---|---|---|
Kymora (core33) |
33 | 3.18 ms | 314,450 | Baseline (1.0×) | 0.0964 µs | Baseline (1.0×) |
catch22 |
22 | 833.1 ms | 1,200 | 262× slower | 37.87 µs | 393× |
TSFEL |
156 | 2,541.6 ms | 393 | 799× slower | 16.29 µs | 169× |
tsfresh |
777 | 20,891.2 ms | 48 | 6,570× slower | 26.89 µs | 279× |
Pooled medians: kymora over 4 HEAD rounds (n=400 runs), competitors over 10 rounds (n=53/83/64); 95% bootstrap CIs in benchmarks/results/F1_REPORT.md. Raw time answers "how long for the batch"; per-feature answers "how expensive each number is". Different feature counts mean different work — read both columns, and prefer the equal-feature table above for a like-for-like claim.
Honest losses, with evidence (benchmarks/results/2026-10-05_l1_rerun/l1_root_cause.json, benchmarks/results/L1_ROOT_CAUSE.md, B3_REPORT.md):
- Tiny single-shot calls. Against a hand-written numba baseline on the same 33 features (HEAD rerun 2026-10-05), kymora loses only at 1 × 100 — gaussian 1.78× [1.76,1.80], heavy_tailed 1.75× [1.74,1.76], random_walk 1.43× [1.42,1.45] (95% CIs exclude 1.0; classified call-overhead + quantile-stage). At 10 × 500 the rows are parity/noise (0.85×–1.08×) and the 1 × 100 sinusoid row is a kymora win with a clamped CI — all four are exploratory (noisy host, CI95 rel-widths 13–78%; the rerun recorded no env/CPU load or per-row CV), so no claim is made off them. Absolute medians at 1 × 100 are ~0.029 ms vs numba 0.016–0.020 ms.
- Thread wake at small n (pre pool-cache, 0.8.0 prep). At 100 × 100, 16 threads made kymora ~27% slower than 1 thread (pool wake/join swamps ~12 µs/series of compute). Scaling efficiency at small batches is poor regardless of core count.
- Single-series tail latency at very short lengths (pre pool-cache, 0.8.0 prep). At length 10, kymora wins the median (345.8 µs vs numpy's 1,317.6 µs) but loses the tail (p99 15.4 ms, max 29.1 ms) to wake jitter.
- Not a robust loss (pre pool-cache, 0.8.0 prep): the 100 × 500 rows against the numba baseline are parity within noise (0.94×–0.99×, CI crosses 1.0).
Guidance: batch many series into one call, or use StreamingExtractor for one-series-at-a-time ingest. At ≥1,000 series the picture inverts — kymora beats the numba baseline 3.4×–11.1× on identical features (equal-feature table above).
- Kymora (
core33, 100,000 × 500): +417.7 MiB peak over a bare interpreter — input 385.4 MiB, extraction overhead +32.3 MiB, output matrix 25.2 MiB (= 100,000 × 33 × 8 B). - tsfresh (EfficientFCParameters, 777 features): extraction peak +281.9 MiB at 1,000 × 500 and +407.5 MiB at 10,000 × 500 (long-format DataFrame plus intermediate frames).
- The tsfresh 100,000 × 500 cell is a multi-hour run — PENDING, tracked in
CLAIMS.md.
Artifact: benchmarks/results/2026-10-05_remeasure/memory.jsonl.
Kymora deliberately computes 33 high-signal, non-redundant features spanning all temporal domains (exact catalog: kymora.feature_names()):
- Moments: mean, std, var, skewness, kurtosis.
- Extrema & quantiles: min, max, median, quantile_10, quantile_25, quantile_75, quantile_90.
- Energy & shape: abs_energy, root_mean_square, number_of_peaks.
- Differences & trend: mean_abs_change, mean_change, cid_ce, mean_second_derivative_central, trend_slope, trend_r2.
- Crossings & runs: zero_crossings, mean_crossings, longest_strike_above_mean, longest_strike_below_mean.
- Autocorrelation: autocorr_lag_1, autocorr_lag_2, autocorr_lag_5, autocorr_lag_10.
- Spectral: dominant_frequency, spectral_centroid, spectral_entropy.
- Complexity: permutation_entropy (order 3, delay 1).
Contributions, bug reports, and PRs are welcome!
Please check CONTRIBUTING.md for details on setting up the local Rust/Python development environment and running the benchmark suites.
Distributed under the MIT License. See LICENSE for details.
Built with Rust and Python by Aamod.