Chronos (Amazon) — AI foundation model za vremenske serije, test na našim svećama (2026-10-01)

↑ Nazad

Zaključak: NE-GO. Ni jedan od tri testirana Chronos modela nema merljiv signal za smer, za magnitudu ni za profit. 0 kandidata u 45 ćelija otkrivačkog skupa po unapred registrovanom kriterijumu. Jedina stvar koju model radi dobro je kalibrisana nesigurnost: interval q10–q90 obuhvata 72–88% ishoda, a cilj je 80%. Ali ni tu nije bolji od trivijalne istorijske volatilnosti (held-out: 3/15 ćelija, a prag je bio 8/15). Ovo je drugi foundation model koji pada na istom testu, posle Kronosa (09-01/09).

Povod

Instagram reel (@jesserurka, 30.9.2026): „$24,000/year terminal vs. a free model on GitHub“, uz „komentariši FORECAST“, dakle levak za skupljanje kontakata. Repo naveden u komentarima je amazon-science/chronos-forecasting. To nije Kronos koji smo već testirali:

	Kronos (testiran 09-01)	Chronos (ovaj test)
Autor	Tsinghua, AAAI 2026	Amazon Science
Šta modeluje	sveće (OHLCV tokeni)	opšte vremenske serije
Veličine	4M / 25M / 102M (499M zatvoren)	Chronos-2 120M, Bolt 9–205M, T5 8–710M
Kovarijate	ne	da (Chronos-2: multivarijantno + kovarijate)
Brzina kod nas (CPU)	8 s/slučaj (small)	0.02–0.06 s/slučaj

Šta je testirano

Oznaka	Model	Ulaz
`c2`	`amazon/chronos-2` (120M)	close
`c2cov`	`amazon/chronos-2` (120M)	close + past kovarijate: log(volume), (high−low)/close
`bolt`	`amazon/chronos-bolt-base` (205M)	close

Metod je isti kao za Kronos, uz dve dopune: PnL i held-out posle datuma izlaska modela.

USDT parovi (66–68 simbola), TF 1h/4h/1d, kontekst 512 zatvorenih sveća, 24 unapred, horizonti h = 1/3/6/12/24. Slučajevi bez rupa (536 uzastopnih sveća), cene > 0.
Otkrivački skup: 240 slučajeva po TF-u (seed 1), cela istorija.
Held-out skup: 240 po TF-u (seed 2), samo origin ≥ 2025-11-01, posle izlaska Chronos-2, dakle sigurno van pretreninga. Origin-i: 2025-11-01 → 2026-09-06.
Baselini: majority (češći smer, upareni McNemar), „predvidi nulu“ (MAE, Wilcoxon), IC (Spearman) i IC uz kontrolu trivijalnog momentuma (parcijalna korelacija).
PnL (lekcija iz Kronosa: tačnost smera ≠ profit): long kad model kaže „gore“, inače flat, provizija 0.1% po strani. Poređenje sa „uvek long“, bootstrap 95% CI.
Kriterijum je zapisan pre prvog rezultata (PREREG): kandidat = (a) hit iznad majority na Bonferroni p < 0.0033, ili (b) IC|mom > 2.94·SE, ili (c) PnL CI iznad „uvek long“. Kandidat prolazi samo ako se ponovi na held-out skupu.

Rezultat: smer, magnituda, PnL

Model	Skup	Smer: bolje od majority	Smer: značajno gore	Magnituda: tuče „nulu“	Kandidati
`c2`	otkrivački	0/15	2/15	0/15	0
`c2cov`	otkrivački	0/15	1/15 (4h/h=24: −14.2 pp, p=0.001)	0/15	0
`bolt`	otkrivački	0/15	0/15	0/15	0
`c2`	held-out	0/15	1/15	0/15	0
`bolt`	held-out	0/15	4/15	0/15	3 (vidi niže)

IC se kreće od −0.135 do +0.124. Ni jedna ćelija ne prelazi Bonferroni prag (2.94·SE ≈ 0.19).
Kovarijate ne pomažu. c2cov je u proseku isto što i c2, a jedina značajna ćelija mu je gora.
Veći model ne pomaže. bolt (205M) nije bolji od c2 (120M). Isto je bilo kod Kronosa: skaliranje 25M → 102M nije donelo ništa.
Model ne ekstrapolira momentum: kontrola na trivijalni momentum menja IC modela za ≤ 0.01.

Tri „PnL“ ćelije Bolt-a na held-out skupu nisu signal

Held-out period (nov 2025 → sep 2026) je pretežno medveđi. Strategija koja je deo vremena flat zato tuče „uvek long“ samo time što manje drži poziciju. Provera protiv nasumičnog signala sa istim udelom long pozicija (5000 permutacija):

Ćelija	udeo long	strategija	nasumično, ista izloženost	p	otkrivački skup, ista ćelija
1h / h=1	42%	−0.078%	−0.084%	0.40	+0.05 (šum)
1h / h=12	58%	−0.018%	−0.236%	0.03	−0.31 (suprotan znak)
4h / h=3	32%	+0.116%	−0.074%	0.04	−0.18 (suprotan znak)

1h/h=1 je čista ušteda na proviziji. Ostale dve su 2 od 30 held-out ćelija na p ≈ 0.03–0.04, što se po slučaju i očekuje (~1.5). Na otkrivačkom skupu imaju suprotan znak. To je isti potpis lažnog nalaza kao dva puta kod Kronosa (granicni-nalaz-replikuj-na-svezem-uzorku). Nisu ni bile kandidati iz otkrivačkog skupa, pa po PREREG-u ne mogu ni da „prođu“.

Jedino „pametno“: kalibrisana nesigurnost, ali bez prednosti

Interval q10–q90 pokriva 72–88% stvarnih ishoda (cilj 80%), pa model dobro zna koliko se cena može pomeriti. Hipoteza H-vol je dopisana u PREREG pre gledanja held-out-a: širina intervala predviđa |prinos| bolje od trivijalne vol = std(log prinosa 48 sveća)·√h.

Model	Otkrivački: značajno bolje	Held-out: značajno bolje	Prag	Ishod
`c2`	2/15	3/15	≥ 8/15	pada
`bolt`	0/15	0/15	≥ 8/15	pada

Oba IC-a su visoka (≈ +0.2 do +0.5), ali jednaka. Model je naučio isto što i jedna linija koda: „posle mirnog perioda mirno, posle burnog burno“. Za stop/TP rastojanja i veličinu pozicije ne donosi ništa što ATR ili rolling std već ne daju, a dodaje zavisnost od 120M modela.

Šta ovo NE isključuje

Nije rađen fine-tuning (repo ga podržava). Uz našu istoriju (Kronos, meta-score, offset sweep, qcv2 overfit) najverovatniji ishod je isti, samo skuplji.
Nisu testirani chronos-t5-large (710M) ni kovarijate iz drugih izvora (funding, OI, BTC kao kovarijata za altove). Ova poslednja je jedina nova ideja koju Chronos-2 omogućava, a Kronos nije. Hipoteza je slaba: ni naši feature-i ni sami generatori nisu pokazali edge.
Testiran je samo „sledeći potez cene“. Model nije testiran kao filter postojećih generatora.

Pouka

Drugi foundation model, isti ishod. Problem nije model nego cilj: kratkoročni smer kripto cene iz same istorije cene nije predvidiv sa ovim alatima, a statistika prognoze (RankIC, CRPS, fev-bench) nije trading edge. Repo i rad to ni ne tvrde; tvrdi to Instagram levak.

Artefakti

Skripte i kriterijum su u repou (Kronos skripte su se izgubile sa scratchpadom): 2026-10-01_chronos/ — PREREG.md (kriterijum pre rezultata), sample_cases.py, infer.py, evaluate.py, voltest.py, eval_*.md. Slučajevi i predikcije (cases*.npz, pred_*.npz) se regenerišu iz sample_cases.py (seed 1/2) i infer.py. Okruženje: Python 3.13, torch 2.14 CPU, chronos-forecasting 2.3.2, Haswell-EP 48 niti.

Dodatak A — pune tabele, otkrivački skup

Legenda: Δpp = hit − majority; p McN = McNemar model vs majority; long% = udeo „gore“ predikcija; IC|mom = IC uz kontrolu momentuma; PnL% = prosečan prinos strategije po slučaju posle provizije; long% (drugi) = „uvek long“; cov80 = % ishoda unutar q10–q90.

c2 (otkrivački)

TF	h	N	hit%	maj%	Δpp	p McN	long%	MAE%	MAE0%	IC	IC mom	IC	mom	PnL%	long%	Δ PnL (95% CI)	cov80
1h	1	240	50.4	53.8	-3.3	0.530	48	0.84	0.76	-0.042	+0.003	-0.042	-0.09	-0.13	+0.04 (-0.06, +0.15)	80	—
1h	3	240	52.9	55.0	-2.1	0.712	51	1.36	1.32	+0.028	-0.075	+0.016	+0.04	+0.03	+0.01 (-0.17, +0.20)	78	—
1h	6	240	52.5	52.9	-0.4	1.000	51	1.87	1.83	+0.020	-0.049	+0.018	+0.06	+0.08	-0.02 (-0.30, +0.26)	75	—
1h	12	240	53.8	50.8	+2.9	0.555	57	2.92	2.86	+0.011	+0.019	+0.011	+0.44	+0.41	+0.03 (-0.27, +0.33)	75	—
1h	24	240	57.1	52.9	+4.2	0.403	52	4.07	4.04	+0.098	-0.008	+0.100	+0.85	+0.91	-0.06 (-0.59, +0.43)	77	—
4h	1	240	49.2	51.7	-2.5	0.640	52	1.68	1.59	-0.008	-0.113	-0.018	+0.01	+0.06	-0.05 (-0.36, +0.19)	79	—
4h	3	240	45.4	50.4	-5.0	0.315	50	2.91	2.86	-0.038	-0.066	-0.041	-0.23	-0.00	-0.22 (-0.84, +0.28)	79	—
4h	6	240	49.2	55.8	-6.7	0.125	40	4.46	4.44	-0.027	-0.107	-0.023	+0.31	+0.33	-0.02 (-1.03, +0.81)	76	—
4h	12	240	46.7	57.5	-10.8	0.017	46	6.18	6.11	-0.008	-0.059	+0.002	+0.15	+0.34	-0.19 (-1.46, +0.80)	76	—
4h	24	240	49.2	57.5	-8.3	0.045	38	7.97	7.78	+0.026	+0.006	+0.025	+0.65	+0.98	-0.33 (-1.58, +0.83)	80	—
1d	1	240	50.4	52.1	-1.7	0.749	63	3.70	3.28	-0.017	-0.039	-0.018	-0.24	-0.29	+0.05 (-0.32, +0.42)	83	—
1d	3	240	52.9	52.1	+0.8	0.933	59	5.71	5.43	-0.003	-0.124	-0.009	+0.01	-0.10	+0.11 (-0.42, +0.63)	81	—
1d	6	240	53.3	52.9	+0.4	1.000	45	8.06	7.62	-0.076	-0.038	-0.072	-0.54	-0.68	+0.14 (-0.81, +1.04)	78	—
1d	12	240	52.9	53.3	-0.4	1.000	42	12.48	12.27	+0.024	+0.058	+0.019	+0.56	-0.23	+0.79 (-0.69, +2.23)	76	—
1d	24	240	52.1	55.0	-2.9	0.500	33	19.21	18.53	+0.010	+0.019	+0.006	+0.54	+1.10	-0.56 (-3.01, +1.72)	72	—

Smer: 0/15 značajno bolje od majority, 2/15 značajno gore (p<0.05, bez korekcije). Magnituda: 0/15 tuče 'predvidi nulu'. Kandidati po PREREG: nema

c2cov (otkrivački)

TF	h	N	hit%	maj%	Δpp	p McN	long%	MAE%	MAE0%	IC	IC mom	IC	mom	PnL%	long%	Δ PnL (95% CI)	cov80
1h	1	240	48.3	53.8	-5.4	0.283	48	0.80	0.76	-0.075	+0.003	-0.075	-0.13	-0.13	+0.00 (-0.11, +0.11)	84	—
1h	3	240	53.3	55.0	-1.7	0.779	52	1.34	1.32	-0.003	-0.075	-0.009	+0.03	+0.03	+0.01 (-0.16, +0.17)	79	—
1h	6	240	52.1	52.9	-0.8	0.930	47	1.81	1.83	+0.081	-0.049	+0.083	+0.08	+0.08	+0.01 (-0.24, +0.25)	79	—
1h	12	240	55.0	50.8	+4.2	0.387	55	2.86	2.86	+0.050	+0.019	+0.049	+0.44	+0.41	+0.03 (-0.25, +0.30)	77	—
1h	24	240	54.2	52.9	+1.2	0.853	51	4.12	4.04	+0.062	-0.008	+0.063	+0.56	+0.91	-0.35 (-1.25, +0.34)	75	—
4h	1	240	47.5	51.7	-4.2	0.395	53	1.68	1.59	-0.031	-0.113	-0.037	-0.02	+0.06	-0.09 (-0.38, +0.16)	81	—
4h	3	240	47.9	50.4	-2.5	0.640	52	2.89	2.86	-0.024	-0.066	-0.026	-0.11	-0.00	-0.11 (-0.71, +0.35)	80	—
4h	6	240	49.2	55.8	-6.7	0.133	42	4.43	4.44	-0.017	-0.107	-0.020	+0.26	+0.33	-0.07 (-1.06, +0.72)	78	—
4h	12	240	48.8	57.5	-8.7	0.071	51	6.09	6.11	+0.005	-0.059	+0.008	+0.10	+0.34	-0.24 (-1.51, +0.80)	75	—
4h	24	240	43.3	57.5	-14.2	0.001	41	7.89	7.78	-0.026	+0.006	-0.028	+0.17	+0.98	-0.81 (-2.57, +0.57)	80	—
1d	1	240	51.7	52.1	-0.4	1.000	62	4.03	3.28	+0.021	-0.039	+0.022	-0.10	-0.29	+0.19 (-0.14, +0.54)	81	—
1d	3	240	53.3	52.1	+1.2	0.867	60	6.01	5.43	+0.050	-0.124	+0.045	+0.24	-0.10	+0.34 (-0.21, +0.91)	80	—
1d	6	240	55.8	52.9	+2.9	0.576	48	8.11	7.62	+0.012	-0.038	+0.015	-0.28	-0.68	+0.41 (-0.50, +1.27)	78	—
1d	12	240	51.7	53.3	-1.7	0.775	46	12.70	12.27	+0.012	+0.058	+0.013	-0.15	-0.23	+0.08 (-1.42, +1.58)	75	—
1d	24	240	52.9	55.0	-2.1	0.620	27	18.96	18.53	+0.056	+0.019	+0.055	+0.93	+1.10	-0.17 (-2.84, +2.43)	73	—

Smer: 0/15 značajno bolje od majority, 1/15 značajno gore (p<0.05, bez korekcije). Magnituda: 0/15 tuče 'predvidi nulu'. Kandidati po PREREG: nema

bolt (otkrivački)

TF	h	N	hit%	maj%	Δpp	p McN	long%	MAE%	MAE0%	IC	IC mom	IC	mom	PnL%	long%	Δ PnL (95% CI)	cov80
1h	1	240	47.5	53.8	-6.2	0.235	42	0.84	0.76	-0.049	+0.003	-0.049	-0.08	-0.13	+0.05 (-0.07, +0.16)	83	—
1h	3	240	48.8	55.0	-6.3	0.203	50	1.38	1.32	-0.020	-0.075	-0.027	-0.06	+0.03	-0.09 (-0.26, +0.07)	79	—
1h	6	240	51.7	52.9	-1.2	0.847	55	1.92	1.83	-0.006	-0.049	-0.012	+0.04	+0.08	-0.04 (-0.24, +0.16)	80	—
1h	12	240	49.2	50.8	-1.7	0.767	57	3.13	2.86	-0.101	+0.019	-0.101	+0.10	+0.41	-0.31 (-0.75, +0.06)	74	—
1h	24	240	45.8	52.9	-7.1	0.111	58	4.59	4.04	-0.135	-0.008	-0.135	+0.19	+0.91	-0.72 (-1.63, -0.03)	69	—
4h	1	240	48.8	51.7	-2.9	0.603	45	1.80	1.59	-0.000	-0.113	-0.002	-0.05	+0.06	-0.11 (-0.44, +0.15)	81	—
4h	3	240	49.6	50.4	-0.8	0.927	51	2.87	2.86	+0.048	-0.066	+0.042	-0.18	-0.00	-0.18 (-0.78, +0.31)	82	—
4h	6	240	52.5	55.8	-3.3	0.539	54	4.45	4.44	+0.071	-0.107	+0.073	-0.16	+0.33	-0.49 (-1.79, +0.53)	80	—
4h	12	240	49.6	57.5	-7.9	0.121	56	6.13	6.11	+0.118	-0.059	+0.114	+0.43	+0.34	+0.09 (-0.92, +0.94)	77	—
4h	24	240	51.7	57.5	-5.8	0.258	55	8.04	7.78	+0.124	+0.006	+0.127	+1.00	+0.98	+0.01 (-1.34, +1.21)	75	—
1d	1	240	48.3	52.1	-3.8	0.515	37	4.38	3.28	-0.014	-0.039	-0.015	-0.21	-0.29	+0.08 (-0.36, +0.52)	88	—
1d	3	240	53.3	52.1	+1.2	0.844	43	6.03	5.43	+0.060	-0.124	+0.059	+0.01	-0.10	+0.11 (-0.64, +0.83)	88	—
1d	6	240	50.8	52.9	-2.1	0.714	50	8.36	7.62	+0.013	-0.038	+0.008	-0.31	-0.68	+0.37 (-0.59, +1.32)	82	—
1d	12	240	48.8	53.3	-4.6	0.367	51	13.36	12.27	-0.003	+0.058	+0.009	-0.42	-0.23	-0.19 (-2.01, +1.45)	79	—
1d	24	240	52.1	55.0	-2.9	0.603	55	20.00	18.53	+0.014	+0.019	+0.019	+0.03	+1.10	-1.07 (-3.76, +1.27)	73	—

Smer: 0/15 značajno bolje od majority, 0/15 značajno gore (p<0.05, bez korekcije). Magnituda: 0/15 tuče 'predvidi nulu'. Kandidati po PREREG: nema

Dodatak B — held-out (origin ≥ 2025-11-01), deskriptivno

c2_holdout

TF	h	N	hit%	maj%	Δpp	p McN	long%	MAE%	MAE0%	IC	IC mom	IC	mom	PnL%	long%	Δ PnL (95% CI)	cov80
1h	1	240	50.8	52.9	-2.1	0.699	55	0.64	0.57	-0.131	-0.053	-0.146	-0.14	-0.20	+0.07 (-0.00, +0.13)	81	—
1h	3	240	49.6	52.9	-3.3	0.519	51	0.94	0.92	-0.042	-0.022	-0.047	-0.09	-0.09	+0.01 (-0.12, +0.13)	82	—
1h	6	240	52.1	52.1	+0.0	1.000	43	1.26	1.24	-0.042	-0.116	-0.048	-0.06	-0.15	+0.09 (-0.07, +0.27)	82	—
1h	12	240	47.5	52.1	-4.6	0.347	47	2.26	2.09	-0.073	-0.097	-0.066	-0.29	-0.40	+0.11 (-0.11, +0.34)	79	—
1h	24	240	47.9	50.4	-2.5	0.617	42	3.45	3.11	-0.089	-0.046	-0.083	-0.19	+0.13	-0.33 (-0.78, +0.09)	75	—
4h	1	240	47.1	54.2	-7.1	0.165	55	1.39	1.33	-0.055	-0.068	-0.060	-0.01	-0.07	+0.06 (-0.07, +0.19)	82	—
4h	3	240	51.2	53.8	-2.5	0.645	51	2.32	2.31	+0.018	-0.081	+0.019	-0.10	-0.24	+0.14 (-0.12, +0.40)	80	—
4h	6	240	50.8	52.9	-2.1	0.731	44	3.29	3.23	+0.014	-0.065	+0.016	-0.20	-0.11	-0.10 (-0.65, +0.43)	78	—
4h	12	240	52.5	53.8	-1.2	0.859	47	4.77	4.68	-0.033	-0.108	-0.028	-0.24	+0.10	-0.34 (-1.06, +0.34)	76	—
4h	24	240	56.7	56.7	+0.0	1.000	38	5.92	5.77	+0.017	-0.007	+0.018	-0.33	-1.10	+0.77 (-0.09, +1.62)	82	—
1d	1	240	49.6	56.2	-6.7	0.236	67	3.51	2.96	-0.011	-0.075	-0.026	-0.47	-0.48	+0.02 (-0.47, +0.42)	81	—
1d	3	240	45.8	54.6	-8.7	0.094	60	5.44	4.89	-0.105	-0.108	-0.136	-0.69	-0.85	+0.16 (-0.48, +0.77)	81	—
1d	6	240	48.3	55.8	-7.5	0.079	39	7.53	6.91	-0.141	-0.037	-0.142	-1.26	-1.07	-0.18 (-1.31, +0.84)	80	—
1d	12	240	49.6	60.8	-11.2	0.007	39	10.53	9.77	-0.180	+0.004	-0.184	-1.64	-1.97	+0.33 (-0.93, +1.58)	78	—
1d	24	240	54.2	60.0	-5.8	0.065	21	15.93	13.94	-0.207	-0.229	-0.138	-2.05	-2.53	+0.48 (-1.62, +2.49)	78	—

Smer: 0/15 značajno bolje od majority, 1/15 značajno gore (p<0.05, bez korekcije). Magnituda: 0/15 tuče 'predvidi nulu'. Kandidati po PREREG: nema

bolt_holdout

TF	h	N	hit%	maj%	Δpp	p McN	long%	MAE%	MAE0%	IC	IC mom	IC	mom	PnL%	long%	Δ PnL (95% CI)	cov80
1h	1	240	49.6	52.9	-3.3	0.554	42	0.65	0.57	+0.014	-0.053	+0.015	-0.08	-0.20	+0.12 (+0.04, +0.21)	84	c
1h	3	240	53.3	52.9	+0.4	1.000	49	0.92	0.92	+0.025	-0.022	+0.025	-0.04	-0.09	+0.05 (-0.09, +0.17)	84	—
1h	6	240	50.0	52.1	-2.1	0.702	55	1.32	1.24	-0.025	-0.116	-0.025	-0.03	-0.15	+0.12 (-0.03, +0.29)	82	—
1h	12	240	48.3	52.1	-3.8	0.498	58	2.18	2.09	+0.025	-0.097	+0.023	-0.02	-0.40	+0.38 (+0.05, +0.78)	77	c
1h	24	240	47.9	50.4	-2.5	0.685	63	3.27	3.11	-0.067	-0.046	-0.073	+0.05	+0.13	-0.09 (-0.62, +0.45)	71	—
4h	1	240	54.2	54.2	+0.0	1.000	24	1.56	1.33	+0.077	-0.068	+0.075	+0.15	-0.07	+0.22 (-0.01, +0.45)	86	—
4h	3	240	53.3	53.8	-0.4	1.000	32	2.45	2.31	+0.028	-0.081	+0.024	+0.12	-0.24	+0.35 (+0.01, +0.71)	82	c
4h	6	240	46.7	52.9	-6.2	0.261	35	3.52	3.23	-0.044	-0.065	-0.048	-0.05	-0.11	+0.06 (-0.47, +0.57)	78	—
4h	12	240	48.3	53.8	-5.4	0.309	42	4.97	4.68	-0.005	-0.108	-0.007	-0.11	+0.10	-0.21 (-0.89, +0.46)	77	—
4h	24	240	47.1	56.7	-9.6	0.033	45	6.18	5.77	-0.009	-0.007	-0.009	-0.68	-1.10	+0.42 (-0.36, +1.17)	80	—
1d	1	240	52.1	56.2	-4.2	0.268	28	4.95	2.96	-0.024	-0.075	-0.030	-0.27	-0.48	+0.21 (-0.36, +0.72)	88	—
1d	3	240	50.0	54.6	-4.6	0.260	33	6.74	4.89	-0.097	-0.108	-0.106	-0.47	-0.85	+0.38 (-0.46, +1.20)	90	—
1d	6	240	45.8	55.8	-10.0	0.028	46	8.71	6.91	-0.092	-0.037	-0.095	-0.54	-1.07	+0.53 (-0.38, +1.42)	86	—
1d	12	240	46.7	60.8	-14.2	0.003	52	11.86	9.77	-0.125	+0.004	-0.125	-1.34	-1.97	+0.62 (-0.44, +1.66)	82	—
1d	24	240	48.3	60.0	-11.7	0.018	54	16.86	13.94	-0.094	-0.229	-0.121	-1.66	-2.53	+0.88 (-0.44, +2.20)	78	—

Smer: 0/15 značajno bolje od majority, 4/15 značajno gore (p<0.05, bez korekcije). Magnituda: 0/15 tuče 'predvidi nulu'. Kandidati po PREREG: [('1h', 1, ['c']), ('1h', 12, ['c']), ('4h', 3, ['c'])]

Dodatak C — H-vol (širina intervala vs trivijalna volatilnost)

H-vol c2 (otkrivački) — 2/15 značajno bolje od trivijalne vol

TF	h	IC širina→	r		IC trivijalna vol
1h	1	+0.439	+0.420	+0.019 (-0.045, +0.080)
1h	3	+0.522	+0.458	+0.063 (+0.005, +0.123)	✔
1h	6	+0.391	+0.339	+0.053 (-0.001, +0.103)
1h	12	+0.413	+0.433	-0.019 (-0.074, +0.037)
1h	24	+0.301	+0.313	-0.012 (-0.077, +0.054)
4h	1	+0.316	+0.379	-0.062 (-0.135, +0.019)
4h	3	+0.381	+0.383	-0.001 (-0.076, +0.081)
4h	6	+0.412	+0.361	+0.051 (-0.023, +0.127)
4h	12	+0.345	+0.353	-0.008 (-0.076, +0.062)
4h	24	+0.225	+0.260	-0.035 (-0.118, +0.041)
1d	1	+0.357	+0.335	+0.022 (-0.074, +0.120)
1d	3	+0.245	+0.298	-0.053 (-0.138, +0.023)
1d	6	+0.186	+0.191	-0.005 (-0.091, +0.084)
1d	12	+0.246	+0.129	+0.117 (+0.042, +0.205)	✔
1d	24	+0.067	+0.027	+0.039 (-0.055, +0.128)

H-vol bolt (otkrivački) — 0/15 značajno bolje od trivijalne vol

TF	h	IC širina→	r
1h	1	+0.444	+0.420	+0.024 (-0.035, +0.087)
1h	3	+0.487	+0.458	+0.029 (-0.031, +0.091)
1h	6	+0.383	+0.339	+0.044 (-0.005, +0.098)
1h	12	+0.443	+0.433	+0.010 (-0.044, +0.064)
1h	24	+0.328	+0.313	+0.015 (-0.043, +0.074)
4h	1	+0.329	+0.379	-0.050 (-0.132, +0.032)
4h	3	+0.356	+0.383	-0.026 (-0.112, +0.067)
4h	6	+0.390	+0.361	+0.029 (-0.047, +0.105)
4h	12	+0.334	+0.353	-0.019 (-0.099, +0.050)
4h	24	+0.252	+0.260	-0.008 (-0.101, +0.079)
1d	1	+0.244	+0.335	-0.091 (-0.212, +0.025)
1d	3	+0.229	+0.298	-0.069 (-0.163, +0.022)
1d	6	+0.141	+0.191	-0.049 (-0.143, +0.047)
1d	12	+0.165	+0.129	+0.035 (-0.056, +0.126)
1d	24	-0.073	+0.027	-0.101 (-0.208, +0.004)

H-vol c2_holdout — 3/15 značajno bolje od trivijalne vol

TF	h	IC širina→	r		IC trivijalna vol
1h	1	+0.412	+0.355	+0.057 (-0.004, +0.122)
1h	3	+0.319	+0.241	+0.078 (+0.020, +0.144)	✔
1h	6	+0.379	+0.355	+0.023 (-0.049, +0.090)
1h	12	+0.398	+0.386	+0.012 (-0.056, +0.084)
1h	24	+0.295	+0.241	+0.053 (-0.025, +0.133)
4h	1	+0.270	+0.289	-0.019 (-0.106, +0.065)
4h	3	+0.365	+0.350	+0.014 (-0.053, +0.081)
4h	6	+0.293	+0.281	+0.013 (-0.056, +0.082)
4h	12	+0.320	+0.366	-0.046 (-0.117, +0.017)
4h	24	+0.399	+0.433	-0.033 (-0.109, +0.041)
1d	1	+0.254	+0.295	-0.040 (-0.124, +0.046)
1d	3	+0.452	+0.337	+0.115 (+0.042, +0.184)	✔
1d	6	+0.444	+0.350	+0.094 (+0.028, +0.165)	✔
1d	12	+0.407	+0.351	+0.056 (-0.015, +0.126)
1d	24	+0.227	+0.220	+0.008 (-0.064, +0.083)

H-vol bolt_holdout — 0/15 značajno bolje od trivijalne vol

TF	h	IC širina→	r
1h	1	+0.362	+0.355	+0.007 (-0.061, +0.081)
1h	3	+0.302	+0.241	+0.061 (-0.004, +0.128)
1h	6	+0.377	+0.355	+0.022 (-0.039, +0.083)
1h	12	+0.437	+0.386	+0.051 (-0.011, +0.114)
1h	24	+0.305	+0.241	+0.063 (-0.013, +0.147)
4h	1	+0.254	+0.289	-0.035 (-0.120, +0.055)
4h	3	+0.378	+0.350	+0.027 (-0.043, +0.099)
4h	6	+0.285	+0.281	+0.004 (-0.065, +0.072)
4h	12	+0.344	+0.366	-0.022 (-0.083, +0.040)
4h	24	+0.424	+0.433	-0.008 (-0.074, +0.059)
1d	1	+0.180	+0.295	-0.115 (-0.235, +0.001)
1d	3	+0.260	+0.337	-0.077 (-0.183, +0.027)
1d	6	+0.304	+0.350	-0.046 (-0.135, +0.047)
1d	12	+0.265	+0.351	-0.087 (-0.179, +0.011)
1d	24	+0.137	+0.220	-0.083 (-0.174, +0.013)

Izvor: 82_2026-10-01_chronos_foundation_model.md