skaters 0.17.0

Rust port of the skaters online distributional forecasting core, parity-gated against the Python reference.
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
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
//! Invertible online transforms for conjugation. Port of src/skaters/transform.py.
//!
//! Each transform has `forward(&mut self, y) -> f64` (runs per observation) and
//! `inverse_k(&self, dists) -> Vec<Dist>` (maps k transformed-space predictions
//! back). Python initialises transform state lazily on the first call; here an
//! `init` flag reproduces the first-call special cases exactly.

use crate::dist::Dist;
use crate::mathx::fsum;
use serde::{Deserialize, Serialize};

#[derive(Clone, Debug, Serialize, Deserialize)]
pub enum Transform {
    Difference(Difference),
    FracDiff(FracDiff),
    Standardize(Standardize),
    EmaT(EmaT),
    Ou(Ou),
    Theta(Theta),
    Drift(Drift),
    Holt(Holt),
    Garch(Garch),
    Seasonal(Seasonal),
    SeasonalAnchor(SeasonalAnchor),
    Power(Power),
    YeoJohnson(YeoJohnson),
    Ar(Ar),
    GroupedAr(GroupedAr),
}

impl Transform {
    pub fn forward(&mut self, y: f64) -> f64 {
        match self {
            Transform::Difference(t) => t.forward(y),
            Transform::FracDiff(t) => t.forward(y),
            Transform::Standardize(t) => t.forward(y),
            Transform::EmaT(t) => t.forward(y),
            Transform::Ou(t) => t.forward(y),
            Transform::Theta(t) => t.forward(y),
            Transform::Drift(t) => t.forward(y),
            Transform::Holt(t) => t.forward(y),
            Transform::Garch(t) => t.forward(y),
            Transform::Seasonal(t) => t.forward(y),
            Transform::SeasonalAnchor(t) => t.forward(y),
            Transform::Power(t) => t.forward(y),
            Transform::YeoJohnson(t) => t.forward(y),
            Transform::Ar(t) => t.forward(y),
            Transform::GroupedAr(t) => t.forward(y),
        }
    }

    pub fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        match self {
            Transform::Difference(t) => t.inverse_k(dists),
            Transform::FracDiff(t) => t.inverse_k(dists),
            Transform::Standardize(t) => t.inverse_k(dists),
            Transform::EmaT(t) => t.inverse_k(dists),
            Transform::Ou(t) => t.inverse_k(dists),
            Transform::Theta(t) => t.inverse_k(dists),
            Transform::Drift(t) => t.inverse_k(dists),
            Transform::Holt(t) => t.inverse_k(dists),
            Transform::Garch(t) => t.inverse_k(dists),
            Transform::Seasonal(t) => t.inverse_k(dists),
            Transform::SeasonalAnchor(t) => t.inverse_k(dists),
            Transform::Power(t) => t.inverse_k(dists),
            Transform::YeoJohnson(t) => t.inverse_k(dists),
            Transform::Ar(t) => t.inverse_k(dists),
            Transform::GroupedAr(t) => t.inverse_k(dists),
        }
    }
}

// ---------------------------------------------------------------------------
// difference
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Difference {
    pub last: f64,
    pub init: bool,
}

pub fn difference() -> Transform {
    Transform::Difference(Difference {
        last: 0.0,
        init: false,
    })
}

impl Difference {
    fn forward(&mut self, y: f64) -> f64 {
        if !self.init {
            self.init = true;
            self.last = y;
            return 0.0;
        }
        let dy = y - self.last;
        self.last = y;
        dy
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let anchor = self.last;
        let mut result = Vec::with_capacity(dists.len());
        let mut cumsum_mean = 0.0;
        let mut cumsum_var = 0.0;
        for d in dists {
            cumsum_mean += d.mean();
            cumsum_var += d.var();
            let std = if cumsum_var > 0.0 {
                cumsum_var.sqrt()
            } else {
                d.std().max(1e-12)
            };
            result.push(Dist::gaussian(anchor + cumsum_mean, std));
        }
        result
    }
}

// ---------------------------------------------------------------------------
// fractional difference
// ---------------------------------------------------------------------------

fn frac_diff_weights(d: f64, window: usize) -> Vec<f64> {
    let mut w = vec![1.0];
    for i in 1..window {
        let last = *w.last().unwrap();
        w.push(-last * (d - i as f64 + 1.0) / i as f64);
    }
    w
}

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct FracDiff {
    pub window: usize,
    pub w_fwd: Vec<f64>,
    pub buffer: Vec<f64>,
}

pub fn fractional_difference(d: f64, window: usize) -> Transform {
    Transform::FracDiff(FracDiff {
        window,
        w_fwd: frac_diff_weights(d, window),
        buffer: Vec::new(),
    })
}

impl FracDiff {
    fn forward(&mut self, y: f64) -> f64 {
        self.buffer.push(y);
        if self.buffer.len() > self.window {
            self.buffer.remove(0);
        }
        let n = self.buffer.len();
        fsum((0..n).map(|j| self.w_fwd[j] * self.buffer[n - 1 - j]))
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        // Recovered horizons feed back through the same weights, so each carries a
        // row of responses to the future innovations and Var = sum_i c_i^2 var_i
        // (issue #255).
        let hn = dists.len();
        let mut buf = self.buffer.clone();
        let mut cbuf: Vec<Option<Vec<f64>>> = vec![None; buf.len()];
        let mut result = Vec::with_capacity(hn);
        for (h, d_in) in dists.iter().enumerate() {
            buf.push(0.0);
            cbuf.push(None);
            let n = buf.len();
            let mut shift = 0.0;
            let mut coef = vec![0.0; hn];
            coef[h] = 1.0;
            for j in 1..n.min(self.window) {
                shift -= self.w_fwd[j] * buf[n - 1 - j];
                if let Some(cj) = &cbuf[n - 1 - j] {
                    for i in 0..h {
                        coef[i] -= self.w_fwd[j] * cj[i];
                    }
                }
            }
            let recovered_mean = d_in.mean() + shift;
            buf[n - 1] = recovered_mean;
            let mut total_var = 0.0;
            for i in 0..=h {
                total_var += coef[i] * coef[i] * dists[i].var();
            }
            cbuf[n - 1] = Some(coef);
            let std = if total_var > 0.0 {
                total_var.sqrt()
            } else {
                d_in.std().max(1e-12)
            };
            result.push(Dist::gaussian(recovered_mean, std));
            if buf.len() > self.window {
                buf.remove(0);
                cbuf.remove(0);
            }
        }
        result
    }
}

// ---------------------------------------------------------------------------
// standardize
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Standardize {
    pub alpha: f64,
    pub eps: f64,
    pub mu: f64,
    pub var: f64,
    pub init: bool,
}

pub fn standardize(alpha: f64, eps: f64) -> Transform {
    Transform::Standardize(Standardize {
        alpha,
        eps,
        mu: 0.0,
        var: 0.0,
        init: false,
    })
}

impl Standardize {
    fn forward(&mut self, y: f64) -> f64 {
        if !self.init {
            self.init = true;
            self.mu = y;
            self.var = 0.0;
            return 0.0;
        }
        let mu = self.mu;
        let diff = y - mu;
        // Emit against the PRIOR state, so the forward map is the affine change
        // of coordinates z = (y - mu) / sigma that the inverse applies. Scaling
        // by the post-update std makes the emission self-normalized (bounded by
        // 1/sqrt(alpha)) and the affine inverse is then not its inverse.
        // Cold start: before the variance is informative, scale by the first
        // nonzero residual, so the first informative emission is +-1.
        let sigma = if self.var > self.eps * self.eps {
            self.var.sqrt()
        } else if diff.abs() > self.eps {
            diff.abs()
        } else {
            self.eps
        };
        let y_prime = diff / sigma;
        self.mu = mu + self.alpha * diff;
        self.var = (1.0 - self.alpha) * self.var + self.alpha * diff * diff;
        y_prime
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let sigma = if self.var > 1e-16 {
            self.var.sqrt()
        } else {
            1e-8
        };
        dists.iter().map(|d| d.affine(sigma, self.mu)).collect()
    }
}

// ---------------------------------------------------------------------------
// ema transform
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct EmaT {
    pub alpha: f64,
    pub level: f64,
    pub init: bool,
}

pub fn ema_transform(alpha: f64) -> Transform {
    assert!(0.0 < alpha && alpha < 1.0);
    Transform::EmaT(EmaT {
        alpha,
        level: 0.0,
        init: false,
    })
}

impl EmaT {
    fn forward(&mut self, y: f64) -> f64 {
        if !self.init {
            self.init = true;
            self.level = y;
            return 0.0;
        }
        let residual = y - self.level;
        self.level += self.alpha * residual;
        residual
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let level = self.level;
        let mut out = Vec::with_capacity(dists.len());
        // alpha^2 * sum_{i<h} var_i: the level's accumulated response (ETS(A,N,N), issue #254)
        let mut add = 0.0;
        for d in dists {
            if add > 0.0 {
                out.push(Dist::new(
                    d.components
                        .iter()
                        .map(|&(w, m, s)| (w, m + level, (s * s + add).sqrt()))
                        .collect(),
                ));
            } else {
                out.push(d.shift(level));
            }
            add += self.alpha * self.alpha * d.var();
        }
        out
    }
}

// ---------------------------------------------------------------------------
// Ornstein-Uhlenbeck transform
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Ou {
    pub alpha: f64,
    pub phi: f64,
    pub m: f64,
    pub fc: f64,
    pub y: f64,
    pub init: bool,
}

pub fn ou_transform(kappa: f64, alpha: f64) -> Transform {
    assert!(0.0 < kappa && kappa <= 1.0);
    assert!(0.0 < alpha && alpha < 1.0);
    Transform::Ou(Ou {
        alpha,
        phi: 1.0 - kappa,
        m: 0.0,
        fc: 0.0,
        y: 0.0,
        init: false,
    })
}

impl Ou {
    fn forward(&mut self, y: f64) -> f64 {
        // Python: `if state is None or not math.isfinite(y)` — a non-finite y
        // RESETS the state even after initialisation.
        if !self.init || !y.is_finite() {
            let y0 = if y.is_finite() { y } else { 0.0 };
            self.init = true;
            self.m = y0;
            self.fc = y0;
            self.y = y0;
            return 0.0;
        }
        let mut resid = y - self.fc;
        if !resid.is_finite() {
            resid = 0.0;
        }
        let m = self.m + self.alpha * (y - self.m);
        let fc = m + self.phi * (y - m);
        self.m = m;
        self.fc = fc;
        self.y = y;
        resid
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let (m, ylast) = (self.m, self.y);
        let phi = self.phi;
        let mut out = Vec::with_capacity(dists.len());
        for (h0, d) in dists.iter().enumerate() {
            let h = (h0 + 1) as f64;
            let center = m + libm::pow(phi, h) * (ylast - m);
            // Var = sum_{i<=h0} phi^{2(h0-i)} var_i, as a scale on this horizon's Dist
            // (issue #256); equals the old sqrt(sum_j phi^2j) for equal variances.
            let v_h = d.var();
            let g = if v_h > 0.0 {
                let mut tot = 0.0;
                for i in 0..=h0 {
                    tot += libm::pow(phi, (2 * (h0 - i)) as f64) * dists[i].var();
                }
                (tot / v_h).sqrt()
            } else if phi < 1.0 - 1e-9 {
                ((1.0 - libm::pow(phi, 2.0 * h)) / (1.0 - phi * phi)).sqrt()
            } else {
                h.sqrt()
            };
            out.push(d.scale(g).shift(center));
        }
        out
    }
}

// ---------------------------------------------------------------------------
// theta
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Theta {
    pub alpha: f64,
    pub ses: f64,
    pub t: u64,
    pub sum_t: f64,
    pub sum_t2: f64,
    pub sum_y: f64,
    pub sum_ty: f64,
    pub slope: f64,
    pub init: bool,
}

pub fn theta(alpha: f64) -> Transform {
    assert!(0.0 < alpha && alpha < 1.0);
    Transform::Theta(Theta {
        alpha,
        ses: 0.0,
        t: 0,
        sum_t: 0.0,
        sum_t2: 0.0,
        sum_y: 0.0,
        sum_ty: 0.0,
        slope: 0.0,
        init: false,
    })
}

impl Theta {
    fn forward(&mut self, y: f64) -> f64 {
        if !self.init {
            self.init = true;
            self.ses = y;
            self.t = 1;
            self.sum_t = 1.0;
            self.sum_t2 = 1.0;
            self.sum_y = y;
            self.sum_ty = y;
            self.slope = 0.0; // Python: key absent, read via .get(..., 0.0)
            return 0.0;
        }
        self.t += 1;
        let t = self.t as f64;

        let forecast = self.ses + self.slope / 2.0;
        let residual = y - forecast;

        self.ses = self.alpha * y + (1.0 - self.alpha) * self.ses;

        self.sum_t += t;
        self.sum_t2 += t * t;
        self.sum_y += y;
        self.sum_ty += t * y;

        let n = t;
        let denom = n * self.sum_t2 - self.sum_t * self.sum_t;
        self.slope = if denom.abs() > 1e-12 {
            (n * self.sum_ty - self.sum_t * self.sum_y) / denom
        } else {
            0.0
        };
        residual
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let mut result = Vec::with_capacity(dists.len());
        for (h, d) in dists.iter().enumerate() {
            let forecast = self.ses + (h + 1) as f64 * self.slope / 2.0 + d.mean();
            // SES: c_0 = 1, c_j = alpha; Var = sum_j c_j^2 var_{h-j} (issue #253).
            let mut total_var = 0.0;
            for j in 0..=h {
                let c = if j == 0 { 1.0 } else { self.alpha };
                total_var += c * c * dists[h - j].var();
            }
            let std = if total_var > 0.0 {
                total_var.sqrt()
            } else {
                d.std().max(1e-12)
            };
            result.push(Dist::gaussian(forecast, std));
        }
        result
    }
}

// ---------------------------------------------------------------------------
// drift
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Drift {
    pub alpha: f64,
    pub decay: f64,
    pub last: f64,
    pub mu: f64,
    pub init: bool,
}

pub fn drift(alpha: f64, shrinkage: f64) -> Transform {
    assert!(0.0 < alpha && alpha < 1.0);
    assert!((0.0..1.0).contains(&shrinkage));
    Transform::Drift(Drift {
        alpha,
        decay: 1.0 - alpha - shrinkage,
        last: 0.0,
        mu: 0.0,
        init: false,
    })
}

impl Drift {
    fn forward(&mut self, y: f64) -> f64 {
        if !self.init {
            self.init = true;
            self.last = y;
            self.mu = 0.0;
            return 0.0;
        }
        let dy = y - self.last;
        let residual = dy - self.mu;
        self.mu = self.decay * self.mu + self.alpha * dy;
        self.last = y;
        residual
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let anchor = self.last;
        let mu = self.mu;
        let rho = self.decay + self.alpha; // 1 - shrinkage: persistence of the drift estimate
        let mut result = Vec::with_capacity(dists.len());
        let mut cumsum_mean = 0.0;
        for (h, d) in dists.iter().enumerate() {
            cumsum_mean += d.mean();
            let total_mean = anchor + (h + 1) as f64 * mu + cumsum_mean;
            // c = 1 + alpha * sum_{m<h-i} rho^m: direct plus through the drift
            // estimate (issue #257).
            let mut total_var = 0.0;
            for i in 0..=h {
                let mut acc = 0.0;
                let mut f = 1.0;
                for _ in 0..(h - i) {
                    acc += f;
                    f *= rho;
                }
                let c = 1.0 + self.alpha * acc;
                total_var += c * c * dists[i].var();
            }
            let total_std = if total_var > 0.0 {
                total_var.sqrt()
            } else {
                d.std().max(1e-12)
            };
            result.push(Dist::gaussian(total_mean, total_std));
        }
        result
    }
}

// ---------------------------------------------------------------------------
// Holt linear
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Holt {
    pub alpha: f64,
    pub beta: f64,
    pub level: f64,
    pub trend: f64,
    pub init: bool,
}

pub fn holt_linear(alpha: f64, beta: f64) -> Transform {
    assert!(0.0 < alpha && alpha < 1.0);
    assert!(0.0 < beta && beta < 1.0);
    Transform::Holt(Holt {
        alpha,
        beta,
        level: 0.0,
        trend: 0.0,
        init: false,
    })
}

impl Holt {
    fn forward(&mut self, y: f64) -> f64 {
        if !self.init {
            self.init = true;
            self.level = y;
            self.trend = 0.0;
            return 0.0;
        }
        let l_prev = self.level;
        let b_prev = self.trend;
        let l_new = self.alpha * y + (1.0 - self.alpha) * (l_prev + b_prev);
        let b_new = self.beta * (l_new - l_prev) + (1.0 - self.beta) * b_prev;
        let residual = y - (l_prev + b_prev);
        self.level = l_new;
        self.trend = b_new;
        residual
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let mut result = Vec::with_capacity(dists.len());
        for (h, d) in dists.iter().enumerate() {
            let forecast = self.level + (h + 1) as f64 * self.trend + d.mean();
            // ETS(A,A,N): c_0 = 1, c_j = alpha + alpha*beta*j; Var = sum_j c_j^2 var_{h-j} (issue #242).
            let mut total_var = 0.0;
            for j in 0..=h {
                let c = if j == 0 { 1.0 } else { self.alpha + self.alpha * self.beta * j as f64 };
                total_var += c * c * dists[h - j].var();
            }
            let std = if total_var > 0.0 {
                total_var.sqrt()
            } else {
                d.std().max(1e-12)
            };
            result.push(Dist::gaussian(forecast, std));
        }
        result
    }
}

// ---------------------------------------------------------------------------
// GARCH(1,1) volatility scaling
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Garch {
    pub omega: f64,
    pub alpha: f64,
    pub beta: f64,
    pub mean_alpha: f64,
    pub eps: f64,
    pub var: f64,
    pub last_dev: f64,
    pub mu: f64,
    pub init: bool,
}

pub fn garch(omega: f64, alpha: f64, beta: f64, mean_alpha: f64, eps: f64) -> Transform {
    assert!(omega > 0.0 && alpha >= 0.0 && beta >= 0.0);
    Transform::Garch(Garch {
        omega,
        alpha,
        beta,
        mean_alpha,
        eps,
        var: 0.0,
        last_dev: 0.0,
        mu: 0.0,
        init: false,
    })
}

pub fn garch_default() -> Transform {
    garch(0.01, 0.1, 0.85, 0.05, 1e-8)
}

impl Garch {
    fn forward(&mut self, y: f64) -> f64 {
        if !self.init {
            self.init = true;
            let persist = self.alpha + self.beta;
            let var0 = if persist < 1.0 {
                self.omega / (1.0 - persist)
            } else {
                self.omega / self.eps
            };
            self.var = var0;
            self.last_dev = 0.0;
            self.mu = y;
            return 0.0;
        }
        // GARCH conditional variance of the DEVIATION from a running mean (not
        // the raw value): shift-invariant, so a level series does not report a
        // volatility of order |y|. On mean-zero returns mu stays ~0 and this
        // reduces to the original y/sigma.
        let dev = y - self.mu;
        let var = self.omega + self.alpha * self.last_dev * self.last_dev + self.beta * self.var;
        let sigma = if var > self.eps * self.eps {
            var.sqrt()
        } else {
            self.eps
        };
        self.var = var;
        self.last_dev = dev;
        self.mu += self.mean_alpha * dev;
        dev / sigma
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let sigma = if self.var > 1e-16 {
            self.var.sqrt()
        } else {
            1e-8
        };
        dists.iter().map(|d| d.affine(sigma, self.mu)).collect()
    }
}

// ---------------------------------------------------------------------------
// seasonal difference
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Seasonal {
    pub period: usize,
    pub buffer: Vec<f64>,
    pub init: bool,
}

pub fn seasonal_difference(period: usize) -> Transform {
    assert!(period >= 1);
    Transform::Seasonal(Seasonal {
        period,
        buffer: Vec::new(),
        init: false,
    })
}

impl Seasonal {
    fn forward(&mut self, y: f64) -> f64 {
        if !self.init {
            self.init = true;
            self.buffer = vec![y];
            return 0.0;
        }
        let y_prime = if self.buffer.len() >= self.period {
            y - self.buffer[self.buffer.len() - self.period]
        } else {
            0.0
        };
        self.buffer.push(y);
        if self.buffer.len() > 2 * self.period {
            self.buffer.remove(0);
        }
        y_prime
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let buf = &self.buffer;
        let period = self.period;
        let mut recovered_means: Vec<f64> = Vec::new();
        let mut recovered_vars: Vec<f64> = Vec::new();
        let mut result = Vec::with_capacity(dists.len());
        for h in 0..dists.len() {
            let (anchor_mean, anchor_var);
            if h < period {
                let buf_idx = buf.len() as i64 - period as i64 + h as i64;
                anchor_mean = if buf_idx >= 0 && (buf_idx as usize) < buf.len() {
                    buf[buf_idx as usize]
                } else {
                    0.0
                };
                anchor_var = 0.0;
            } else {
                let lag_idx = h - period;
                anchor_mean = recovered_means[lag_idx];
                anchor_var = recovered_vars[lag_idx];
            }
            recovered_means.push(dists[h].mean() + anchor_mean);
            recovered_vars.push(dists[h].var() + anchor_var);
            if anchor_var > 0.0 {
                result.push(Dist::new(
                    dists[h]
                        .components
                        .iter()
                        .map(|&(w, m, s)| (w, m + anchor_mean, (s * s + anchor_var).sqrt()))
                        .collect(),
                ));
            } else {
                result.push(dists[h].shift(anchor_mean));
            }
        }
        result
    }
}


// ---------------------------------------------------------------------------
// seasonal anchor: hedged seasonal location (phase-EMA blended with naive)
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct SeasonalAnchor {
    pub period: usize,
    pub alpha: f64,
    pub weight: f64,
    pub ema: Vec<Option<f64>>,
    pub buffer: Vec<f64>,
    pub n: usize,
    pub init: bool,
}

pub fn seasonal_anchor(period: usize, alpha: f64, weight: f64) -> Transform {
    assert!(period >= 1);
    assert!(0.0 < alpha && alpha < 1.0);
    assert!((0.0..=1.0).contains(&weight));
    Transform::SeasonalAnchor(SeasonalAnchor {
        period,
        alpha,
        weight,
        ema: vec![None; period],
        buffer: Vec::new(),
        n: 0,
        init: false,
    })
}

impl SeasonalAnchor {
    fn anchor_of(&self, ema: Option<f64>, snaive: f64) -> f64 {
        match ema {
            Some(e) => self.weight * e + (1.0 - self.weight) * snaive,
            None => snaive,
        }
    }

    fn forward(&mut self, y: f64) -> f64 {
        if !self.init {
            self.init = true;
            self.buffer = vec![y];
            self.n = 1;
            return 0.0;
        }
        let p = self.n % self.period;
        let snaive = if self.buffer.len() >= self.period {
            self.buffer[self.buffer.len() - self.period]
        } else {
            *self.buffer.last().unwrap()
        };
        let y_prime = y - self.anchor_of(self.ema[p], snaive);
        self.ema[p] = Some(match self.ema[p] {
            Some(e) => e + self.alpha * (y - e),
            None => y,
        });
        self.buffer.push(y);
        if self.buffer.len() > 2 * self.period {
            self.buffer.remove(0);
        }
        self.n += 1;
        y_prime
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let buf = &self.buffer;
        let period = self.period;
        let hn = dists.len();
        let mut recovered_means: Vec<f64> = Vec::new();
        // Innovation-response rows of recovered values and of each phase's EMA: the
        // forward updates the EMA from recovered values too, so an anchor at
        // h >= period responds to the value one period back through (1 - weight)
        // and weight * alpha (issue #258).
        let mut rec_resp: Vec<Vec<f64>> = Vec::new();
        let mut ema_resp: Vec<Vec<f64>> = vec![vec![0.0; hn]; period];
        let mut result = Vec::with_capacity(hn);
        for h in 0..hn {
            let p = (self.n + h) % period;
            let (snaive, sn_resp): (f64, Option<Vec<f64>>);
            if h < period {
                let buf_idx = buf.len() as i64 - period as i64 + h as i64;
                snaive = if buf_idx >= 0 && (buf_idx as usize) < buf.len() {
                    buf[buf_idx as usize]
                } else {
                    *buf.last().unwrap_or(&0.0)
                };
                sn_resp = None;
            } else {
                let lag_idx = h - period;
                snaive = recovered_means[lag_idx];
                sn_resp = Some(rec_resp[lag_idx].clone());
            }
            let e = self.ema[p];
            let a_mean = self.anchor_of(e, snaive);
            let mut a_resp = vec![0.0; hn];
            for i in 0..h {
                let sn = match &sn_resp {
                    Some(v) => v[i],
                    None => 0.0,
                };
                a_resp[i] = match e {
                    Some(_) => self.weight * ema_resp[p][i] + (1.0 - self.weight) * sn,
                    None => sn,
                };
            }
            let mut a_var = 0.0;
            for i in 0..h {
                a_var += a_resp[i] * a_resp[i] * dists[i].var();
            }
            let mut r = a_resp.clone();
            r[h] = 1.0;
            recovered_means.push(dists[h].mean() + a_mean);
            rec_resp.push(r.clone());
            ema_resp[p] = match e {
                None => r.clone(),
                Some(_) => (0..hn)
                    .map(|i| (1.0 - self.alpha) * ema_resp[p][i] + self.alpha * r[i])
                    .collect(),
            };
            if a_var > 0.0 {
                result.push(Dist::new(
                    dists[h]
                        .components
                        .iter()
                        .map(|&(w, m, s)| (w, m + a_mean, (s * s + a_var).sqrt()))
                        .collect(),
                ));
            } else {
                result.push(dists[h].shift(a_mean));
            }
        }
        result
    }
}

// ---------------------------------------------------------------------------
// signed power transform
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Power {
    pub p: f64,
    pub inv_p: f64,
}

pub fn power_transform(p: f64) -> Transform {
    assert!(0.0 < p && p < 1.0);
    Transform::Power(Power { p, inv_p: 1.0 / p })
}

impl Power {
    fn forward(&mut self, y: f64) -> f64 {
        libm::copysign(libm::pow(y.abs(), self.p), y)
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let mut result = Vec::with_capacity(dists.len());
        for d in dists {
            let mut components = Vec::with_capacity(d.components.len());
            for &(w, mu, sigma) in &d.components {
                let orig_mean = libm::copysign(libm::pow(mu.abs(), self.inv_p), mu);
                let abs_mu = mu.abs();
                let deriv = if abs_mu > 1e-12 {
                    self.inv_p * libm::pow(abs_mu, self.inv_p - 1.0)
                } else {
                    self.inv_p
                };
                let orig_std = (sigma * deriv).max(1e-12);
                components.push((w, orig_mean, orig_std));
            }
            result.push(Dist::new(components));
        }
        result
    }
}

// ---------------------------------------------------------------------------
// Yeo-Johnson
// ---------------------------------------------------------------------------

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct YeoJohnson {
    pub l: f64,
}

pub fn yeo_johnson(lmbda: f64) -> Transform {
    Transform::YeoJohnson(YeoJohnson { l: lmbda })
}

impl YeoJohnson {
    fn fwd(&self, y: f64) -> f64 {
        let l = self.l;
        if y >= 0.0 {
            if l == 0.0 {
                libm::log1p(y)
            } else {
                (libm::pow(y + 1.0, l) - 1.0) / l
            }
        } else if l == 2.0 {
            -libm::log1p(-y)
        } else {
            -((libm::pow(-y + 1.0, 2.0 - l) - 1.0) / (2.0 - l))
        }
    }

    fn inv(&self, yp: f64) -> f64 {
        let l = self.l;
        if yp >= 0.0 {
            if l == 0.0 {
                return libm::expm1(yp.min(350.0));
            }
            let base = (l * yp + 1.0).max(1e-12);
            return libm::pow(base, 1.0 / l) - 1.0;
        }
        if l == 2.0 {
            return 1.0 - libm::exp((-yp).min(350.0));
        }
        let base = (-(2.0 - l) * yp + 1.0).max(1e-12);
        1.0 - libm::pow(base, 1.0 / (2.0 - l))
    }

    fn dinv(&self, yp: f64) -> f64 {
        let l = self.l;
        if yp >= 0.0 {
            if l == 0.0 {
                return libm::exp(yp.min(350.0));
            }
            let base = (l * yp + 1.0).max(1e-12);
            return libm::pow(base, 1.0 / l - 1.0);
        }
        if l == 2.0 {
            return libm::exp((-yp).min(350.0));
        }
        let base = (-(2.0 - l) * yp + 1.0).max(1e-12);
        libm::pow(base, 1.0 / (2.0 - l) - 1.0)
    }

    fn forward(&mut self, y: f64) -> f64 {
        self.fwd(y)
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        let mut result = Vec::with_capacity(dists.len());
        for d in dists {
            let comps = d
                .components
                .iter()
                .map(|&(w, mu, sigma)| (w, self.inv(mu), (sigma * self.dinv(mu)).max(1e-12)))
                .collect();
            result.push(Dist::new(comps));
        }
        result
    }
}

// ---------------------------------------------------------------------------
// AR(p) with recursive least squares
// ---------------------------------------------------------------------------

fn mat_vec(m: &[f64], v: &[f64], n: usize) -> Vec<f64> {
    let mut result = vec![0.0; n];
    for i in 0..n {
        let mut s = 0.0;
        for j in 0..n {
            s += m[i * n + j] * v[j];
        }
        result[i] = s;
    }
    result
}

fn dot(a: &[f64], b: &[f64], n: usize) -> f64 {
    fsum((0..n).map(|i| a[i] * b[i]))
}

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct Ar {
    pub p: usize,
    pub lam: f64,
    pub ridge: f64,
    pub decay: f64,
    pub buffer: Vec<f64>,
    pub phi: Vec<f64>,
    /// Stationarity-projected coefficients, re-derived after every update.
    /// Used for BOTH the emitted residual and `inverse_k`'s reconstruction,
    /// so the two always agree on which predictor produced the residual
    /// (skaters#232) -- `phi` alone drives the raw RLS fit. Absent from
    /// checkpoints written before skaters#235: deserialises empty and is
    /// derived from `phi` on first use.
    #[serde(default)]
    pub phi_used: Vec<f64>,
    pub pmat: Vec<f64>,
    pub n: u64,
}

pub fn ar(order: usize, lam: f64, ridge: f64, decay: f64) -> Transform {
    assert!(order >= 1);
    assert!(0.0 < lam && lam <= 1.0);
    assert!(decay >= 0.0);
    let mut t = Ar {
        p: order,
        lam,
        ridge,
        decay,
        buffer: Vec::new(),
        phi: vec![0.0; order],
        phi_used: vec![0.0; order],
        pmat: Vec::new(),
        n: 0,
    };
    t.pmat = t.init_p();
    Transform::Ar(t)
}

pub fn ar_default(order: usize) -> Transform {
    ar(order, 0.99, 1.0, 0.0)
}

impl Ar {
    fn init_p(&self) -> Vec<f64> {
        let p = self.p;
        let mut pm = vec![0.0; p * p];
        for j in 0..p {
            pm[j * p + j] = if self.decay > 0.0 {
                self.ridge / libm::pow((j + 1) as f64, self.decay)
            } else {
                self.ridge
            };
        }
        pm
    }

    fn forward(&mut self, y: f64) -> f64 {
        let p = self.p;
        self.n += 1;
        let residual;
        if self.buffer.len() >= p {
            let bl = self.buffer.len();
            let x: Vec<f64> = (0..p).map(|i| self.buffer[bl - 1 - i]).collect();

            // Emitted residual uses `phi_used` -- the stationarity-projected
            // coefficients as of the end of the previous tick (derived now for
            // a pre-#235 checkpoint, which carries only `phi`).
            if self.phi_used.len() != p {
                self.phi_used = ar_stationary(&self.phi);
            }
            let prediction = fsum((0..p).map(|i| self.phi_used[i] * x[i]));
            residual = y - prediction;

            // RLS keeps fitting the RAW (unconstrained) least-squares problem
            // with its own prediction error.
            let raw_prediction = fsum((0..p).map(|i| self.phi[i] * x[i]));
            let raw_residual = y - raw_prediction;
            let px = mat_vec(&self.pmat, &x, p);
            let denom = self.lam + dot(&x, &px, p);
            if denom.abs() > 1e-15 {
                let kg: Vec<f64> = px.iter().map(|v| v / denom).collect();
                for i in 0..p {
                    self.phi[i] += kg[i] * raw_residual;
                }
                for i in 0..p {
                    for j in 0..p {
                        self.pmat[i * p + j] = (self.pmat[i * p + j] - kg[i] * px[j]) / self.lam;
                    }
                }
                let bad = !self.pmat.iter().all(|v| v.is_finite())
                    || self.pmat.iter().fold(0.0_f64, |a, v| a.max(v.abs())) > 1e10;
                if bad {
                    self.pmat = self.init_p();
                }
            }
            // Re-project after every update so the next forward() call and
            // inverse_k always agree on the same constrained coefficients.
            self.phi_used = ar_stationary(&self.phi);
        } else {
            residual = y;
        }
        self.buffer.push(y);
        if self.buffer.len() > 2 * p + 10 {
            self.buffer.remove(0);
        }
        residual
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        if self.phi_used.len() != self.p {
            // pre-#235 checkpoint: the cache follows from the raw coefficients
            return ar_inverse(dists, &self.buffer, &ar_stationary(&self.phi), self.p);
        }
        ar_inverse(dists, &self.buffer, &self.phi_used, self.p)
    }
}

/// Spectral radius of the AR companion matrix (largest |characteristic root|):
/// the h-step forecast converges iff this is < 1. Closed form for p<=2, power
/// iteration otherwise.
fn ar_spectral_radius(phi: &[f64]) -> f64 {
    let p = phi.len();
    if p == 1 {
        return phi[0].abs();
    }
    if p == 2 {
        let (a, b) = (phi[0], phi[1]);
        let disc = a * a + 4.0 * b;
        if disc >= 0.0 {
            let r = disc.sqrt();
            return ((a + r) / 2.0).abs().max(((a - r) / 2.0).abs());
        }
        // libm::hypot, not f64::hypot: the std method calls the platform's own
        // libm (MSVC's differs from glibc's and Apple's) and is not guaranteed
        // bit-identical, which broke the cross-platform digest on Windows while
        // Linux, macOS and wasm agreed. Every other transcendental here already
        // routes through the libm crate for exactly this reason. Python's
        // math.hypot is also platform libm, so the 1e-6 parity tolerance covers
        // the reference comparison; bit-identity is a Rust-to-Rust claim.
        return libm::hypot(a / 2.0, (-disc).sqrt() / 2.0);
    }
    // Power iteration from a single seed can silently miss the true dominant
    // eigenvalue if that seed happens to be (near-)orthogonal to it. Two
    // seeds unlikely to BOTH align with a non-dominant eigenvector; take the
    // max (skaters#232 comment: phi=[3,-2,0,0,0] seeded at all-ones converges
    // to 1 instead of the true radius 2).
    let mut best = 0.0_f64;
    for seed in [
        vec![1.0_f64; p],
        (0..p)
            .map(|i| if i % 2 == 0 { 1.0 } else { -1.0 })
            .collect::<Vec<f64>>(),
    ] {
        let mut v = seed;
        let mut rho = 0.0;
        for _ in 0..60 {
            let mut nv = vec![0.0_f64; p];
            let mut s = 0.0;
            for j in 0..p {
                s += phi[j] * v[j];
            }
            nv[0] = s;
            for i in 1..p {
                nv[i] = v[i - 1];
            }
            let mut m = nv.iter().fold(0.0_f64, |acc, x| acc.max(x.abs()));
            if m == 0.0 {
                m = 1.0;
            }
            for x in nv.iter_mut() {
                *x /= m;
            }
            v = nv;
            rho = m;
        }
        best = best.max(rho);
    }
    best
}

/// Damp AR coefficients into the stationary region for forecasting. Scaling
/// phi_j by gamma^(j+1) scales every companion eigenvalue by gamma; a no-op when
/// already stationary.
fn ar_stationary(phi: &[f64]) -> Vec<f64> {
    let margin = 0.999;
    let rho = ar_spectral_radius(phi);
    if rho <= margin {
        return phi.to_vec();
    }
    let g = margin / rho;
    // libm::pow, not f64::powi: powi lowers to the llvm.powi intrinsic, whose
    // expansion is target dependent, so it is not bit-stable across platforms.
    // pow also matches the reference more closely, since Python's `g ** (j + 1)`
    // calls C pow rather than doing repeated multiplication.
    (0..phi.len())
        .map(|j| phi[j] * libm::pow(g, (j + 1) as f64))
        .collect()
}

/// Shared AR-style inverse (used by Ar and GroupedAr). `phi` must already be
/// the stationarity-projected coefficients the caller's forward() emitted
/// residuals under (skaters#232) -- not re-derived here from the raw fit.
fn ar_inverse(dists: &[Dist], buf: &[f64], phi: &[f64], p: usize) -> Vec<Dist> {
    let n = dists.len();
    // Impulse responses psi_0..psi_{n-1}: psi_0 = 1, psi_i = sum_j phi_j psi_{i-j}.
    let mut psi = vec![0.0_f64; n];
    if n > 0 {
        psi[0] = 1.0;
    }
    for i in 1..n {
        let mut s = 0.0;
        for j in 0..p {
            if i as i64 - 1 - j as i64 >= 0 {
                s += phi[j] * psi[i - 1 - j];
            }
        }
        psi[i] = s;
    }
    let mut recovered_means: Vec<f64> = Vec::new();
    let mut result = Vec::with_capacity(n);
    for h in 0..n {
        let mut ar_mean = 0.0;
        for j in 0..p {
            let lag_h = h as i64 - j as i64 - 1;
            if lag_h < 0 {
                let buf_idx = buf.len() as i64 + lag_h;
                if buf_idx >= 0 && (buf_idx as usize) < buf.len() {
                    ar_mean += phi[j] * buf[buf_idx as usize];
                }
            } else if (lag_h as usize) < recovered_means.len() {
                ar_mean += phi[j] * recovered_means[lag_h as usize];
            }
        }
        let total_mean = dists[h].mean() + ar_mean;
        // Var(y_{t+h+1}) = sum_i psi_{h-i}^2 var_i (issue #245); same order as Python/JS.
        let mut total_var = 0.0;
        for i in 0..=h {
            total_var += psi[h - i] * psi[h - i] * dists[i].var();
        }
        let total_std = if total_var > 0.0 {
            total_var.sqrt()
        } else {
            dists[h].std().max(1e-12)
        };
        recovered_means.push(total_mean);
        result.push(Dist::gaussian(total_mean, total_std));
    }
    result
}

// ---------------------------------------------------------------------------
// grouped AR
// ---------------------------------------------------------------------------

fn build_groups(max_lag: usize) -> Vec<usize> {
    let mut groups = Vec::new();
    let mut g = 0usize;
    let mut size = 1usize;
    let mut assigned = 0usize;
    while assigned < max_lag {
        for _ in 0..size {
            if assigned >= max_lag {
                break;
            }
            groups.push(g);
            assigned += 1;
        }
        g += 1;
        size *= 2;
    }
    groups
}

#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct GroupedAr {
    pub max_lag: usize,
    pub lam: f64,
    pub ridge: f64,
    pub groups: Vec<usize>,
    pub n_groups: usize,
    pub buffer: Vec<f64>,
    pub theta: Vec<f64>,
    /// Expanded, stationarity-projected per-lag coefficients, re-derived
    /// after every update (see `Ar::phi_used`). Absent from checkpoints
    /// written before skaters#235: deserialises empty, derived from `theta`.
    #[serde(default)]
    pub phi_used: Vec<f64>,
    pub pmat: Vec<f64>,
    pub n: u64,
}

pub fn grouped_ar(max_lag: usize, lam: f64, ridge: f64) -> Transform {
    assert!(max_lag >= 1);
    assert!(0.0 < lam && lam <= 1.0);
    assert!(ridge > 0.0);
    let groups = build_groups(max_lag);
    let n_groups = groups.iter().copied().max().unwrap() + 1;
    let mut pmat = vec![0.0; n_groups * n_groups];
    for i in 0..n_groups {
        pmat[i * n_groups + i] = ridge;
    }
    Transform::GroupedAr(GroupedAr {
        max_lag,
        lam,
        ridge,
        groups,
        n_groups,
        buffer: Vec::new(),
        theta: vec![0.0; n_groups],
        phi_used: vec![0.0; max_lag],
        pmat,
        n: 0,
    })
}

impl GroupedAr {
    fn forward(&mut self, y: f64) -> f64 {
        let ng = self.n_groups;
        self.n += 1;
        let residual;
        if self.buffer.len() >= self.max_lag {
            let bl = self.buffer.len();
            let mut x = vec![0.0; ng];
            for j in 0..self.max_lag {
                x[self.groups[j]] += self.buffer[bl - 1 - j];
            }

            // Emitted residual uses the previous tick's stationarity-projected,
            // per-lag coefficients -- same fix as Ar (skaters#232).
            let lag_x: Vec<f64> = (0..self.max_lag).map(|j| self.buffer[bl - 1 - j]).collect();
            if self.phi_used.len() != self.max_lag {
                let raw: Vec<f64> = (0..self.max_lag).map(|j| self.theta[self.groups[j]]).collect();
                self.phi_used = ar_stationary(&raw);
            }
            let prediction = fsum((0..self.max_lag).map(|j| self.phi_used[j] * lag_x[j]));
            residual = y - prediction;

            // RLS keeps fitting the RAW group-level least-squares problem
            // with its own prediction error.
            let raw_prediction = fsum((0..ng).map(|g| self.theta[g] * x[g]));
            let raw_residual = y - raw_prediction;
            let px = mat_vec(&self.pmat, &x, ng);
            let denom = self.lam + dot(&x, &px, ng);
            if denom.abs() > 1e-15 {
                let kg: Vec<f64> = px.iter().map(|v| v / denom).collect();
                for g in 0..ng {
                    self.theta[g] += kg[g] * raw_residual;
                }
                for i in 0..ng {
                    for j in 0..ng {
                        self.pmat[i * ng + j] = (self.pmat[i * ng + j] - kg[i] * px[j]) / self.lam;
                    }
                }
                let bad = !self.pmat.iter().all(|v| v.is_finite())
                    || self.pmat.iter().fold(0.0_f64, |a, v| a.max(v.abs())) > 1e10;
                if bad {
                    self.pmat = vec![0.0; ng * ng];
                    for i in 0..ng {
                        self.pmat[i * ng + i] = self.ridge;
                    }
                }
            }
            let raw_phi: Vec<f64> = (0..self.max_lag).map(|j| self.theta[self.groups[j]]).collect();
            self.phi_used = ar_stationary(&raw_phi);
        } else {
            residual = y;
        }
        self.buffer.push(y);
        if self.buffer.len() > self.max_lag + 10 {
            self.buffer.remove(0);
        }
        residual
    }

    fn inverse_k(&self, dists: &[Dist]) -> Vec<Dist> {
        // Same coefficients forward() already emitted residuals under
        // (skaters#232) -- maintained incrementally there, not re-derived here,
        // except for a pre-#235 checkpoint that has no cache yet.
        if self.phi_used.len() != self.max_lag {
            let raw: Vec<f64> = (0..self.max_lag).map(|j| self.theta[self.groups[j]]).collect();
            return ar_inverse(dists, &self.buffer, &ar_stationary(&raw), self.max_lag);
        }
        ar_inverse(dists, &self.buffer, &self.phi_used, self.max_lag)
    }
}