solow-tsa 0.2.0

Time-series analysis (acf/pacf/adfuller, AR models) for the Solow statistical library
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
//! Seasonal-Trend decomposition using LOESS (STL).
//!
//! A faithful port of the reference `STL(endog, period).fit()` (the Cleveland
//! et al. 1990 algorithm) for the non-robust default configuration. The window
//! lengths follow the reference defaults: `seasonal = 7`, `trend` derived from
//! the period and seasonal window, `low_pass` the smallest odd integer greater
//! than the period; all LOESS degrees are 1 and all jumps are 1. With
//! `robust = false` the fit performs a single outer iteration with five inner
//! iterations and no robustness weights.
//!
//! The loops below deliberately mirror the reference's 1-based, index-driven
//! Fortran/Cython kernel (the loess weighting uses the running index directly),
//! so a handful of index-based clippy lints are silenced module-wide.
#![allow(clippy::needless_range_loop)]
#![allow(clippy::manual_memcpy)]
#![allow(clippy::manual_div_ceil)]

use ndarray::Array1;
use solow_core::error::{Error, Result};

/// The three components produced by [`Stl::fit`].
#[derive(Debug, Clone)]
pub struct StlResult {
    /// Estimated trend component.
    pub trend: Array1<f64>,
    /// Estimated seasonal component.
    pub seasonal: Array1<f64>,
    /// Residual component (`endog - trend - seasonal`).
    pub resid: Array1<f64>,
}

/// Seasonal-Trend decomposition using LOESS.
#[derive(Debug, Clone)]
pub struct Stl {
    endog: Vec<f64>,
    period: usize,
    seasonal: usize,
    trend: usize,
    low_pass: usize,
    seasonal_deg: i32,
    trend_deg: i32,
    low_pass_deg: i32,
    seasonal_jump: usize,
    trend_jump: usize,
    low_pass_jump: usize,
}

fn make_odd(mut v: usize) -> usize {
    if v % 2 == 0 {
        v += 1;
    }
    v
}

impl Stl {
    /// Create an STL model for `endog` with the given seasonal `period`, using
    /// the reference default windows and degrees.
    pub fn new(endog: Array1<f64>, period: usize) -> Result<Self> {
        if period < 2 {
            return Err(Error::Value("period must be at least 2".into()));
        }
        let n = endog.len();
        if n < 2 * period {
            return Err(Error::Value(
                "endog must have at least two full periods".into(),
            ));
        }
        solow_core::tools::ensure_all_finite(&endog.view(), "endog")?;
        let seasonal = 7usize; // reference default, odd >= 3
                               // trend = ceil(1.5 * period / (1 - 1.5 / seasonal)) made odd.
        let trend_f = 1.5 * period as f64 / (1.0 - 1.5 / seasonal as f64);
        let trend = make_odd(trend_f.ceil() as usize);
        // low_pass = smallest odd integer > period.
        let low_pass = make_odd(period + 1);
        Ok(Self {
            endog: endog.to_vec(),
            period,
            seasonal,
            trend,
            low_pass,
            seasonal_deg: 1,
            trend_deg: 1,
            low_pass_deg: 1,
            seasonal_jump: 1,
            trend_jump: 1,
            low_pass_jump: 1,
        })
    }

    /// Override the seasonal smoother length (must be odd and `>= 3`).
    pub fn with_seasonal(mut self, seasonal: usize) -> Self {
        self.seasonal = seasonal;
        self
    }

    /// Override the trend smoother length.
    pub fn with_trend(mut self, trend: usize) -> Self {
        self.trend = trend;
        self
    }

    /// Override the low-pass smoother length.
    pub fn with_low_pass(mut self, low_pass: usize) -> Self {
        self.low_pass = low_pass;
        self
    }

    /// Fit the decomposition (non-robust: 5 inner iterations, no outer loop).
    pub fn fit(&self) -> Result<StlResult> {
        let mut inner = StlInner::new(self);
        inner.run(5);
        let n = self.endog.len();
        let mut resid = Array1::<f64>::zeros(n);
        for i in 0..n {
            resid[i] = self.endog[i] - inner.season[i] - inner.trend[i];
        }
        Ok(StlResult {
            trend: Array1::from(inner.trend.clone()),
            seasonal: Array1::from(inner.season.clone()),
            resid,
        })
    }
}

/// Working state mirroring the reference Cython implementation. Arrays use the
/// same layout: `work[0..=4]` plus the seasonal working buffer.
struct StlInner<'a> {
    cfg: &'a Stl,
    nobs: usize,
    trend: Vec<f64>,
    season: Vec<f64>,
    rw: Vec<f64>,
    // work[5][nobs + 2*period]
    work: [Vec<f64>; 5],
    season_buf: Vec<f64>, // _season scratch used by _est weights
    use_rw: bool,
}

impl<'a> StlInner<'a> {
    fn new(cfg: &'a Stl) -> Self {
        let nobs = cfg.endog.len();
        let np = cfg.period;
        let len = nobs + 2 * np;
        StlInner {
            cfg,
            nobs,
            trend: vec![0.0; nobs],
            season: vec![0.0; nobs],
            rw: vec![1.0; nobs],
            work: [
                vec![0.0; len],
                vec![0.0; len],
                vec![0.0; len],
                vec![0.0; len],
                vec![0.0; len],
            ],
            season_buf: vec![0.0; len],
            use_rw: false,
        }
    }

    fn run(&mut self, inner_iter: usize) {
        // Non-robust: outer_iter = 0, so one onestep call then stop.
        self.use_rw = false;
        for v in self.trend.iter_mut() {
            *v = 0.0;
        }
        for v in self.season.iter_mut() {
            *v = 0.0;
        }
        for v in self.rw.iter_mut() {
            *v = 1.0;
        }
        self.onestp(inner_iter);
    }

    fn onestp(&mut self, inner_iter: usize) {
        let n = self.nobs;
        let np = self.cfg.period;
        for _ in 0..inner_iter {
            // Detrend.
            for i in 0..n {
                self.work[0][i] = self.cfg.endog[i] - self.trend[i];
            }
            // Seasonal smoothing -> writes work[1] (length n + 2*np).
            self.ss();
            // Low-pass filter of work[1] -> work[2].
            self.fts();
            // LOESS of work[2] (length n) with low_pass window -> work[0].
            {
                let src = self.work[2].clone();
                let mut ys = vec![0.0; n];
                let mut res = vec![0.0; self.work[3].len()];
                let rw = self.work[3].clone();
                self.ess(
                    &src,
                    n,
                    self.cfg.low_pass,
                    self.cfg.low_pass_deg,
                    self.cfg.low_pass_jump,
                    false,
                    &rw,
                    &mut ys,
                    &mut res,
                );
                for i in 0..n {
                    self.work[0][i] = ys[i];
                }
            }
            // Deseasonalize: season = work[1][np + i] - work[0][i].
            for i in 0..n {
                self.season[i] = self.work[1][np + i] - self.work[0][i];
                self.work[0][i] = self.cfg.endog[i] - self.season[i];
            }
            // Trend smoothing -> trend.
            {
                let src = self.work[0].clone();
                let mut ys = self.trend.clone();
                let mut res = self.work[2].clone();
                let rw = self.rw.clone();
                self.ess(
                    &src,
                    n,
                    self.cfg.trend,
                    self.cfg.trend_deg,
                    self.cfg.trend_jump,
                    self.use_rw,
                    &rw,
                    &mut ys,
                    &mut res,
                );
                self.trend = ys;
            }
        }
    }

    /// Seasonal smoothing of cycle-subseries (reference `_ss`).
    fn ss(&mut self) {
        let n = self.nobs;
        let np = self.cfg.period;
        let ns = self.cfg.seasonal;
        let isdeg = self.cfg.seasonal_deg;
        let nsjump = self.cfg.seasonal_jump;
        let userw = self.use_rw;

        // y = work[0]; season output = work[1].
        let y = self.work[0].clone();
        let rw = self.rw.clone();

        let mut work1 = vec![0.0; n + 2 * np];
        let mut work2 = vec![0.0; n + 2 * np];
        let mut work3 = vec![0.0; n + 2 * np];
        let mut work4 = self.season_buf.clone();

        for j in 0..np {
            let k = (n - (j + 1)) / np + 1;
            for i in 0..k {
                work1[i] = y[i * np + j];
            }
            if userw {
                for i in 0..k {
                    work3[i] = rw[i * np + j];
                }
            }
            // LOESS smooth subseries into work2[1..].
            {
                let src = work1.clone();
                let mut ys = vec![0.0; k.max(1)];
                let mut res = work4.clone();
                self.ess(&src, k, ns, isdeg, nsjump, userw, &work3, &mut ys, &mut res);
                for (i, &v) in ys.iter().enumerate().take(k) {
                    work2[1 + i] = v;
                }
                work4 = res;
            }
            // Extend by one point before.
            let xs = 0i64;
            let nright = ns.min(k);
            let before = self.est(
                &work1,
                k,
                ns,
                isdeg,
                xs,
                1,
                nright as i64,
                &mut work4,
                userw,
                &work3,
            );
            work2[0] = if before.is_nan() { work2[1] } else { before };
            // Extend by one point after.
            let xs = (k + 1) as i64;
            let nleft = if k >= ns { (k - ns + 1).max(1) } else { 1 };
            let after = self.est(
                &work1,
                k,
                ns,
                isdeg,
                xs,
                nleft as i64,
                k as i64,
                &mut work4,
                userw,
                &work3,
            );
            work2[k + 1] = if after.is_nan() { work2[k] } else { after };
            // Store into season output (work[1]).
            for m in 0..(k + 2) {
                self.work[1][m * np + j] = work2[m];
            }
        }
        self.season_buf = work4;
    }

    /// Low-pass filter (reference `_fts`): three moving averages.
    fn fts(&mut self) {
        let np = self.cfg.period;
        let n = self.nobs + 2 * np;
        // x = work[1], trend = work[2], work = work[0].
        let x = self.work[1].clone();
        let mut trend = vec![0.0; self.work[2].len()];
        let mut work = vec![0.0; self.work[0].len()];
        ma(&x, n, np, &mut trend);
        let tmp = trend.clone();
        ma(&tmp, n - np + 1, np, &mut work);
        let tmp2 = work.clone();
        ma(&tmp2, n - 2 * np + 2, 3, &mut trend);
        self.work[2] = trend;
    }

    /// LOESS smoothing over `n` points (reference `_ess`).
    #[allow(clippy::too_many_arguments)]
    fn ess(
        &self,
        y: &[f64],
        n: usize,
        len_: usize,
        ideg: i32,
        njump: usize,
        userw: bool,
        rw: &[f64],
        ys: &mut [f64],
        res: &mut [f64],
    ) {
        if n < 2 {
            ys[0] = y[0];
            return;
        }
        let newnj = njump.min(n - 1);
        if len_ >= n {
            let nleft = 1i64;
            let nright = n as i64;
            let mut i = 0usize;
            while i < n {
                ys[i] = self.est(
                    y,
                    n,
                    len_,
                    ideg,
                    (i + 1) as i64,
                    nleft,
                    nright,
                    res,
                    userw,
                    rw,
                );
                if ys[i].is_nan() {
                    ys[i] = y[i];
                }
                i += newnj;
            }
        } else if newnj == 1 {
            let nsh = (len_ + 2) / 2;
            let mut nleft = 1i64;
            let mut nright = len_ as i64;
            for i in 0..n {
                if (i + 1) > nsh && nright != n as i64 {
                    nleft += 1;
                    nright += 1;
                }
                ys[i] = self.est(
                    y,
                    n,
                    len_,
                    ideg,
                    (i + 1) as i64,
                    nleft,
                    nright,
                    res,
                    userw,
                    rw,
                );
                if ys[i].is_nan() {
                    ys[i] = y[i];
                }
            }
        } else {
            let nsh = (len_ + 1) / 2;
            let mut nleft;
            let mut nright;
            let mut i = 0usize;
            while i < n {
                if (i + 1) < nsh {
                    nleft = 1i64;
                    nright = len_ as i64;
                } else if (i + 1) >= (n - nsh + 1) {
                    nleft = (n - len_ + 1) as i64;
                    nright = n as i64;
                } else {
                    nleft = (i + 1 - nsh + 1) as i64;
                    nright = (len_ + i + 1 - nsh) as i64;
                }
                ys[i] = self.est(
                    y,
                    n,
                    len_,
                    ideg,
                    (i + 1) as i64,
                    nleft,
                    nright,
                    res,
                    userw,
                    rw,
                );
                if ys[i].is_nan() {
                    ys[i] = y[i];
                }
                i += newnj;
            }
        }

        if newnj == 1 {
            return;
        }

        // Linear interpolation between jumped points.
        let mut i = 0usize;
        while i < n.saturating_sub(newnj) {
            let delta = (ys[i + newnj] - ys[i]) / newnj as f64;
            for j in i..(i + newnj) {
                ys[j] = ys[i] + delta * ((j + 1) as f64 - (i + 1) as f64);
            }
            i += newnj;
        }
        // Fill final segment.
        let k = ((n - 1) / newnj) * newnj + 1;
        if k != n {
            // recompute the final point with the last nleft/nright.
            let (nleft, nright) = last_window(n, len_, newnj);
            ys[n - 1] = self.est(y, n, len_, ideg, n as i64, nleft, nright, res, userw, rw);
            if ys[n - 1].is_nan() {
                ys[n - 1] = y[n - 1];
            }
            if k != (n - 1) {
                let delta = (ys[n - 1] - ys[k - 1]) / (n - k) as f64;
                for j in k..n {
                    ys[j] = ys[k - 1] + delta * ((j + 1) as f64 - k as f64);
                }
            }
        }
    }

    /// LOESS local estimate at position `xs` (reference `_est`).
    #[allow(clippy::too_many_arguments)]
    fn est(
        &self,
        y: &[f64],
        n: usize,
        len_: usize,
        ideg: i32,
        xs: i64,
        nleft: i64,
        nright: i64,
        w: &mut [f64],
        userw: bool,
        rw: &[f64],
    ) -> f64 {
        let rng = n as f64 - 1.0;
        let mut h = ((xs - nleft).max(nright - xs)) as f64;
        if len_ > n {
            h += ((len_ - n) / 2) as f64;
        }
        let h9 = 0.999 * h;
        let h1 = 0.001 * h;
        let mut a = 0.0;
        let lo = (nleft - 1) as usize;
        let hi = nright as usize; // exclusive
        for j in lo..hi {
            w[j] = 0.0;
            let r = ((j as i64 + 1 - xs).abs()) as f64;
            if r <= h9 {
                if r <= h1 {
                    w[j] = 1.0;
                } else {
                    w[j] = (1.0 - (r / h).powi(3)).powi(3);
                }
                if userw {
                    w[j] *= rw[j];
                }
                a += w[j];
            }
        }
        if a <= 0.0 {
            return f64::NAN;
        }
        for j in lo..hi {
            w[j] /= a;
        }
        if h > 0.0 && ideg > 0 {
            let mut amean = 0.0;
            for j in lo..hi {
                amean += w[j] * (j as f64 + 1.0);
            }
            let mut b = xs as f64 - amean;
            let mut c = 0.0;
            for j in lo..hi {
                c += w[j] * (j as f64 + 1.0 - amean).powi(2);
            }
            if c.sqrt() > 0.001 * rng {
                b /= c;
                for j in lo..hi {
                    w[j] *= b * (j as f64 + 1.0 - amean) + 1.0;
                }
            }
        }
        let mut ys = 0.0;
        for j in lo..hi {
            ys += w[j] * y[j];
        }
        ys
    }
}

/// Reproduce the `nleft`/`nright` of the last interior `_est` call so the final
/// segment fill matches the reference (which reuses the loop's final window).
fn last_window(n: usize, len_: usize, newnj: usize) -> (i64, i64) {
    let nsh = (len_ + 1) / 2;
    // Last index visited by the jumping loop strictly below n.
    let mut last_i = 0usize;
    let mut i = 0usize;
    while i < n {
        last_i = i;
        i += newnj;
    }
    let ii = last_i + 1; // 1-based
    if ii < nsh {
        (1, len_ as i64)
    } else if ii >= (n - nsh + 1) {
        ((n - len_ + 1) as i64, n as i64)
    } else {
        ((ii - nsh + 1) as i64, (len_ + ii - nsh) as i64)
    }
}

/// Moving average of length `len_` over the first `n` points (reference `_ma`).
fn ma(x: &[f64], n: usize, len_: usize, ave: &mut [f64]) {
    let newn = n - len_ + 1;
    let flen = len_ as f64;
    let mut v = 0.0;
    for &xi in x.iter().take(len_) {
        v += xi;
    }
    ave[0] = v / flen;
    // Slide the length-`len_` window forward one step at a time: the entering
    // sample is `x[m + len_]` and the leaving sample is `x[m]`.
    for (m, ave_j) in ave.iter_mut().take(newn).enumerate().skip(1) {
        v += x[(m - 1) + len_] - x[m - 1];
        *ave_j = v / flen;
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use ndarray::Array1;

    #[test]
    fn default_windows_match_reference_rules() {
        let y = Array1::from_shape_fn(48, |i| i as f64);
        let stl = Stl::new(y, 12).unwrap();
        assert_eq!(stl.seasonal, 7);
        assert_eq!(stl.low_pass, 13);
        // trend = ceil(1.5*12/(1-1.5/7)) -> 23.
        assert_eq!(stl.trend, 23);
    }

    #[test]
    fn components_sum_to_series() {
        let n = 60;
        let pattern = [2.0, -1.0, -3.0, 2.0];
        let y = Array1::from_shape_fn(n, |i| 10.0 + 0.3 * i as f64 + pattern[i % 4]);
        let stl = Stl::new(y.clone(), 4).unwrap();
        let r = stl.fit().unwrap();
        for i in 0..n {
            let recon = r.trend[i] + r.seasonal[i] + r.resid[i];
            assert!((recon - y[i]).abs() < 1e-9, "reconstruction at {i}");
        }
    }

    #[test]
    fn rejects_short_series() {
        let y = Array1::from_vec(vec![1.0, 2.0, 3.0]);
        assert!(Stl::new(y, 4).is_err());
    }
}