Feat/tsfresh parity extractors

.github/workflows/CI.yml

@@ -48,7 +48,7 @@ jobs:
           HYPOTHESIS_PROFILE: ci
         run: |
           # bit-exact parity is checked against these versions
-          uv pip install --system hypothesis tsfresh==0.21.2 numpy==2.4.4 pandas==3.0.2 scipy==1.17.1
+          uv pip install --system hypothesis tsfresh==0.21.2 numpy==2.4.4 pandas==3.0.2 scipy==1.17.1 statsmodels==0.15.0
           pytest -m fuzz
 
   linux:
@@ -75,7 +75,12 @@ jobs:
           CFLAGS_aarch64_unknown_linux_gnu: "-D__ARM_ARCH=8"
         with:
           target: ${{ matrix.platform.target }}
-          args: --release --out dist --find-interpreter
+          # One abi3 wheel for CPython >= 3.11, plus version-specific wheels
+          # for the interpreters abi3 does not cover (free-threaded CPython,
+          # PyPy). --find-interpreter builds only the abi3 wheel. maturin
+          # knows these interpreters' configuration, so they need not be
+          # installed in the build image
+          args: --release --out dist -i python3.11 -i python3.14t -i pypy3.11
           sccache: false
           manylinux: auto
       - name: Upload wheels
@@ -108,7 +113,8 @@ jobs:
         uses: PyO3/maturin-action@v1
         with:
           target: ${{ matrix.platform.target }}
-          args: --release --out dist --find-interpreter
+          # As for manylinux: abi3 plus free-threaded and PyPy wheels
+          args: --release --out dist -i python3.11 -i python3.14t -i pypy3.11
           sccache: false
           manylinux: musllinux_1_2
       - name: Upload wheels

CHANGELOG.md

@@ -33,6 +33,15 @@ All notable changes to TSFX are documented here.
 
 ### Changed
 
+- **Fair TSFresh benchmark.** `python_benchmarks/tsfresh_compare.py`
+  replaces `minimal_bench.py`, `efficient_bench.py`, `1brc/benchmark.py` and
+  `efficient.json`. TSFresh now runs exactly the calculators and parameters
+  TSFX runs for the tier (derived from `.tsfx-config.toml` and checked by
+  feature count), both libraries get the same pre-sorted in-memory data with
+  CSV parsing outside the timer, TSFresh's `n_jobs` matches the Polars
+  thread pool, and both get a warm-up and repeated runs. Previously TSFX's
+  time included reading the CSV, the stocks series were not sorted by date
+  for TSFX, and the feature sets differed.
 - **Many more output columns by default.** The `Efficient` tier now gives
   about 750 columns per value column (was about 155), 400 of them from
   `fft_coefficient`. Existing columns keep their names, but the new ones
@@ -63,6 +72,52 @@ All notable changes to TSFX are documented here.
   scaled to the series or fit) rather than bit for bit, since tsfx does not
   reproduce numpy's FFT, LAPACK and SIMD `log` arithmetic.
 
+### Fixed
+
+- **Linux wheels for free-threaded CPython 3.14 and PyPy 3.11 again.** 0.5.0
+  shipped only the abi3 wheels, which those interpreters cannot use, so
+  installing there built tsfx from source (with a Rust toolchain).
+- **`ar_coefficient` matches statsmodels 0.15.** A series of exactly
+  2k + 1 values (as many observations as AR(k) parameters) now gives the
+  exact fit, as tsfresh does with statsmodels 0.15, instead of NaN (or 0 for
+  `coeff = k`), which matched statsmodels 0.14. For example, a single value
+  with `k = 0` gives that value. The tsfresh fuzz tests in CI now pin
+  `statsmodels==0.15.0`.
+
+### Performance
+
+- **`number_cwt_peaks` is much faster on long series,** with the same
+  results bit for bit: on 50M rows of 1BRC data (series of about 5,600
+  values) it takes 2.6 s instead of 10.9 s. Its noise estimate
+  now selects the 10th percentile of each window instead of sorting it, and
+  ridge lines are matched by binary search instead of comparing every pair.
+  The `Efficient` tier on 50M rows of 1BRC data takes 13.0 s instead of
+  21.9 s.
+- **`augmented_dickey_fuller`, `permutation_entropy` and
+  `change_quantiles` are faster,** with the same results bit for bit.
+  Together they make the `Efficient` tier on 100M rows of 1BRC data take
+  25.5 s instead of 35.3 s (−28%), the features alone measured as follows:
+  - `augmented_dickey_fuller` 16.6 → about 10 s: the check for an exact fit
+    first compares against a cheap upper bound of the fit's scale, and only
+    computes the exact scale (a pass over every row and regressor) when the
+    residual is small enough to matter.
+  - `permutation_entropy` 3.7 → 2.1 s: each window's pattern is counted as
+    one integer instead of an allocated list.
+  - `change_quantiles` 3.0 → 2.5 s: each corridor's changes are computed
+    once and shared by its aggregations.
+- **`augmented_dickey_fuller` and `ar_coefficient` solve well-conditioned
+  regressions from the Gram matrix.** A Householder QR sweeps the whole
+  design once per regressor and was limited by memory bandwidth on long
+  series; the Gram matrix needs one pass. On 100M rows of 1BRC data,
+  `augmented_dickey_fuller` alone takes 4.4 s instead of 12.6 s. **Not bit
+  for bit:** results agree with the previous ones to within about
+  1e-10 (ADF test statistics and p-values within 1e-12 relative on
+  1BRC and edge-case series, the same chosen lags, AR coefficients within
+  2e-11 of the fit's largest coefficient). Ill-conditioned,
+  near-exact and exactly fitting regressions still use the QR decomposition,
+  so their results, including the deviations from tsfresh documented for
+  exact fits, are unchanged.
+
 ## 0.5.0 (2026-10-12)
 
 ### Changed

Cargo.toml

@@ -15,8 +15,8 @@ crate-type = ["lib", "cdylib"]
 serde = { version = "1", features = ["derive"] }
 toml = "1.1"
 # abi3: one wheel per platform for every CPython >= 3.11, including future
-# versions (free-threaded builds, which have no stable ABI, still get
-# version-specific wheels)
+# versions. Free-threaded CPython and PyPy have no stable ABI: on Linux, CI
+# builds version-specific wheels for them (see the -i flags in CI.yml)
 pyo3 = { version = "0.29", features = ["abi3-py311"] }
 chrono = "0.4.42"
 csv = "1.3"

python_benchmarks/1brc/README.md

@@ -1,6 +1,6 @@
 # 1 Billion Row Benchmark
 
-To do perform the benchmark you must first generate data (if you are in the root
+To perform the benchmark you must first generate data (if you are in the root
 directory):
 
 ```bash
@@ -9,14 +9,26 @@ python python_benchmarks/1brc/generate_data.py N
 
 where `N` is a number (1_000_000_000 for the proper 1 billion row challenge).
 
-Then run benchmark like so:
+Then compare TSFX against TSFresh like so:
 
 ```bash
-python python_benchmarks/1brc/benchmark.py
+python python_benchmarks/tsfresh_compare.py --dataset 1brc --rows 10_000_000
 ```
 
-If `tsfx` runs out of memory you can specify how many threads polars should use,
-by setting the environment variable:
+`--rows` limits how many measurements are read (default 10 million, `0` for
+all). For scale, on an 80-core machine with the `efficient` tier (8,859
+stations, best of two timed runs):
+
+| rows | TSFX   | TSFresh |
+| ---- | ------ | ------- |
+| 10M  | 1.3 s  | 52.3 s  |
+| 50M  | 7.1 s  | 192.7 s |
+
+The full billion rows has not been measured. See
+the docstring of `tsfresh_compare.py` for how the comparison is kept fair.
+
+Both libraries use as many threads (TSFresh: processes) as Polars does. To
+limit them, for example if you run out of memory, set:
 
 ```bash
 export POLARS_MAX_THREADS=16