Cascade (SN91) · current king vs its init model

King uid 124 · hotkey 5EpwFZ4iMFwWrdYJ… · receipts index updated 2026-10-03 04:09 UTC · page built 2026-10-03 04:30 UTC · source: public validator receipts

+0.88%

The king's trained model is 0.88% better than the starting (init) checkpoint on the live eval windows (geomean of WQL and MASE; 90% interval +0.38% … +1.40%). This is the whole edge that 5 hours of training on the king's data adds.

Settlements scored

same checkpoint across settlements unless retrained

Eval windows

5,994

64 / 256 / 720 steps, multichannel merged per window

Biggest advantage

+6.7%

IXP traffic

Biggest weakness

-7.9%

kraken price+volume+trades (3-ch)

better than init, confidentbetter, within noise worse, within noiseworse than init, confident · = fewer than 30 windows

Advantage and weakness by source type

Gain of the trained model over the init model on each kind of real series. Right = training helped, left = training made it worse. Hover a bar for details.

Table view

group	init	trained	gain	90% interval	verdict	windows trained better	windows
IXP traffic	0.2684	0.2505	+6.69%	+3.80% … +9.61%	Better than init (confident)	61%	169
CDC NSSP ED visits	0.6202	0.5886	+5.10%	+0.51% … +11.31%	Better than init (confident)	65%	155
COVID / RKI / flu / epi	0.4658	0.4436	+4.77%	-1.07% … +7.42%	Within noise	68%	93
power load / demand	0.2252	0.2179	+3.24%	+1.91% … +5.48%	Better than init (confident)	62%	139
power price / imbalance	0.2441	0.2383	+2.40%	+0.78% … +3.78%	Better than init (confident)	62%	336
downloads / pageviews / Q&A	0.1907	0.1868	+2.04%	+1.42% … +2.45%	Better than init (confident)	53%	193
road/rail/ped traffic counts	0.4445	0.4359	+1.94%	+0.36% … +3.86%	Better than init (confident)	51%	181
bike/dock station status	0.2678	0.2626	+1.93%	+0.92% … +3.05%	Better than init (confident)	51%	693
misskey / social activity	0.5342	0.5245	+1.82%	-0.38% … +5.07%	Within noise	47%	485
311 / permits / licences / registrations	0.5915	0.5828	+1.47%	-0.36% … +4.01%	Within noise	41%	123
crypto price (1-ch / OHLCV)	0.1969	0.1941	+1.43%	+0.57% … +2.53%	Better than init (confident)	57%	268
tide / sea level / waves	0.2043	0.2020	+1.15%	-5.96% … +4.26%	Within noise	52%	89
other	0.3719	0.3687	+0.86%	+0.31% … +1.46%	Better than init (confident)	55%	692
air quality (NO2/PM/O3)	0.5900	0.5873	+0.46%	-0.26% … +1.24%	Within noise	50%	328
911/fire/police/EMS dispatch	0.4445	0.4427	+0.40%	-1.40% … +1.88%	Within noise	54%	126
Tor relays	2.7792	2.7717	+0.27%	+0.27% … +0.27%	Better than init (confident)	43%	21
USDA meat / commodity prices	0.4911	0.4918	-0.15%	-1.00% … +1.06%	Within noise	46%	130
daily FX	0.2734	0.2744	-0.36%	-1.66% … +1.62%	Within noise	46%	127
generation (solar/wind/fuel)	0.3661	0.3675	-0.37%	-1.58% … +0.84%	Within noise	50%	233
policy rates / yields / t-bills	0.4158	0.4184	-0.63%	-2.61% … +0.97%	Within noise	50%	344
BGP / net / ops metrics	0.2064	0.2080	-0.74%	-3.14% … +1.55%	Within noise	48%	64
weather station	0.3208	0.3237	-0.90%	-2.64% … +0.40%	Within noise	47%	123
openFDA / registries	0.5160	0.5221	-1.19%	-3.40% … +0.63%	Within noise	56%	36
river / hydro / groundwater	0.1786	0.1810	-1.30%	-3.49% … +0.57%	Within noise	51%	275
rain	0.5121	0.5203	-1.60%	-1.85% … +0.59%	Within noise	56%	16
retail / product	0.6140	0.6241	-1.65%	-1.92% … +0.35%	Within noise	42%	66
grid flows / CO2 / outages	0.3317	0.3377	-1.81%	-3.58% … -0.36%	Worse than init (confident)	47%	327
kraken price+volume+trades (3-ch)	0.8447	0.9115	-7.91%	-19.63% … +0.07%	Within noise	56%	162

By domain

The validator samples about the same number of windows from each domain.

Table view

group	init	trained	gain	90% interval	verdict	windows trained better	windows
sales	0.4390	0.4293	+2.20%	+0.43% … +4.47%	Better than init (confident)	52%	817
healthcare	0.5661	0.5555	+1.86%	+0.48% … +3.34%	Better than init (confident)	58%	792
web_cloudops	0.3417	0.3359	+1.71%	+0.87% … +2.63%	Better than init (confident)	49%	773
transport	0.3003	0.2957	+1.55%	+0.65% … +2.46%	Better than init (confident)	49%	861
nature	0.2223	0.2208	+0.64%	-0.36% … +1.57%	Within noise	54%	823
energy	0.3392	0.3395	-0.10%	-0.96% … +0.66%	Within noise	53%	855
unknown	0.2759	0.2762	-0.12%	-1.45% … +1.25%	Within noise	52%	254
econ_fin	0.3757	0.3808	-1.36%	-3.97% … +0.38%	Within noise	53%	819

By forecast length

One third of the windows at each length.

Table view

group	init	trained	gain	90% interval	verdict	windows trained better	windows
h64	0.2073	0.2045	+1.33%	+0.57% … +2.11%	Better than init (confident)	52%	1998
h256	0.4041	0.4023	+0.45%	-0.17% … +1.04%	Within noise	52%	1998
h720	0.5132	0.5088	+0.86%	-0.15% … +1.77%	Within noise	54%	1998

Individual sources: largest gains and losses

Sources with at least 8 windows; top 12 and bottom 12.

Table view

group	init	trained	gain	90% interval	verdict	windows trained better	windows
steam_top_games_concurrent_players	0.2892	0.2162	+25.25%	+25.25% … +25.25%	Better than init (confident)	82%	34
cdc_nssp_ed_ari_daily_by_state	0.4338	0.3456	+20.34%	+20.34% … +20.34%	Better than init (confident)	100%	30
gbfs_cat_ganxeta_reus	0.0902	0.0803	+10.99%	+10.99% … +10.99%	Better than init (confident)	67%	9
gbfs_santiago_station_status	0.1049	0.0935	+10.87%	+10.87% … +10.87%	Better than init (confident)	70%	20
bis_policy_rates_daily_panel_a	0.2102	0.1885	+10.35%	+10.35% … +10.35%	Better than init (confident)	58%	12
gbfs_optymo_optymo_belfort_auto_libre_service_belfor	0.1185	0.1093	+7.76%	+7.76% … +7.76%	Better than init (confident)	50%	8
gbfs_publicbikesystem_hibike_kona_hi	0.0309	0.0287	+7.22%	+7.22% … +7.22%	Better than init (confident)	62%	8
marine_ie_tide_gauges	0.2361	0.2191	+7.21%	+7.21% … +7.21%	Better than init (confident)	54%	37
gbfs_buenos_aires_station_status	0.0260	0.0244	+6.20%	+6.20% … +6.20%	Better than init (confident)	36%	11
rki_germany_covid_hospitalizations	0.4355	0.4129	+5.20%	+5.20% … +5.20%	Better than init (confident)	67%	73
gbfs_toronto_station_status	0.1463	0.1389	+5.03%	+5.03% … +5.03%	Better than init (confident)	64%	22
ecb_euro_area_yield_curve_daily_panel	0.6212	0.5907	+4.90%	+4.90% … +4.90%	Better than init (confident)	67%	18
gbfs_trondheim_station_status	0.0664	0.0678	-2.09%	-2.09% … -2.09%	Worse than init (confident)	36%	11
datalakes_eawag_lake_obs	0.0466	0.0477	-2.44%	-2.44% … -2.44%	Worse than init (confident)	33%	24
carbonintensity_gb_genmix	0.5905	0.6055	-2.54%	-2.54% … -2.54%	Worse than init (confident)	53%	17
norges_bank_nok_fx_daily_panel	0.3373	0.3479	-3.13%	-3.13% … -3.13%	Worse than init (confident)	39%	33
gbfs_wienmobil_vienna_station_status	0.4457	0.4618	-3.62%	-3.62% … -3.62%	Worse than init (confident)	29%	17
ipma_pt_surface_observations	1.0669	1.1223	-5.19%	-5.19% … -5.19%	Worse than init (confident)	41%	17
opensensemap_pm10_network	0.4017	0.4241	-5.55%	-5.55% … -5.55%	Worse than init (confident)	56%	9
gbfs_rio_de_janeiro_station_status	0.5410	0.5758	-6.44%	-6.44% … -6.44%	Worse than init (confident)	21%	14
apnic_labs_ipv6_adoption	0.1623	0.1750	-7.82%	-7.82% … -7.82%	Worse than init (confident)	23%	13
statnett_physical_flow	0.2375	0.2586	-8.91%	-8.91% … -8.91%	Worse than init (confident)	42%	31
bis_policy_rates_daily_panel_b	0.1155	0.1267	-9.75%	-9.75% … -9.75%	Worse than init (confident)	50%	10
usace_cwms_timeseries	0.0384	0.0465	-21.16%	-21.16% … -21.16%	Worse than init (confident)	25%	12

Settlements behind these numbers

Each row is one real duel where the king's trained checkpoint and the init model were scored on the same windows by the validators.

settlement block	king role	eval pool	trained checkpoint	init checkpoint	GPU	init	trained	gain	windows	published (UTC)
9192600	challenger	9183600	`sha256:409699fa98ec`	`sha256:3d82fddcaf36`	NVIDIA H100 80GB HBM3	0.3605	0.3567	+1.07%	1998	2026-10-02 04:09
9194400	king	9194400	`sha256:409699fa98ec`	`sha256:3d82fddcaf36`	NVIDIA H100 80GB HBM3	0.3293	0.3292	+0.05%	1998	2026-10-02 10:08
9199800	king	9198000	`sha256:641b773816a9`	`sha256:3da09c722d77`	NVIDIA H100 80GB HBM3	0.3634	0.3579	+1.51%	1998	2026-10-03 04:06

How it is computed: per-window scores from the receipts, channels merged per window and aggregated with the validator's own code (collapse_channels_by_window, global_geomean); intervals are the validator's paired cluster bootstrap (windows from one feed resample together). Domains come from the published pool metadata; "unknown" means that pool's metadata is not published yet (24 h lag).