How this is built

Anyone can publish a dashboard. What is hard to copy is the method underneath — what gets measured, what gets thrown away, and what happens when we are wrong. Every figure on this page is read live from the systems it describes.

What runs, every day Grade D · descriptiveas of 2026-10-19

ReadData

signal modules

130

terminal panels

healthy today

the rest serve last-good, flagged stale

tracked series

46,305

stored daily since 2026-09-04

Counted from the repository and the live build at page-generation time, not maintained by hand — if a module is deleted this number falls.

scalars

field	value
signal modules	130
panels	74
healthy	75
tracked series	46,305
history values	1,385,475
history days	44
history from	2026-09-04
history to	2026-10-18

The calibration study — 121.7M resolved trades Grade C · counted distributionas of 2026-09-06

ReadHistoryData

trades scored

121,671,888

trade-time Brier

0.1685

vs 0.25 coin-flip

calibration error

0.7%

ECE

· Across 121,671,888 resolved trades (both sides scored), Kalshi prices are remarkably well calibrated at trade time: ECE 0.7%, trade-time Brier 0.1685 vs the 0.25 coin-flip baseline.

· Longshot bias is real and asymmetric: takers buying 1-20c contracts average -1.3c EV per contract on the YES side and -1.1c on the NO side — the lottery-ticket buyer funds the market.

· The mirror image: takers of 81-99c favorites average -0.7c (YES) and -0.2c (NO) per contract — being boring pays roughly its keep.

· Longshots still pull volume: 1-20c contracts took 30.5% of taker contract volume in the latest quarter.

Both sides of every trade are scored (the taker at their price, the counterparty at its complement) against the market's actual resolution. The widely-quoted 'day before resolution' Brier of ~0.05 flatters a market by scoring it once the answer is nearly known; trade-time scoring is the honest number, and it is what we hold ourselves to as well. Full study →

The leaf is how well the Kalshi crowd is calibrated, measured over resolved trades. Its own multi-year record exists and is drawn below.

The pane's own leaf over 51 months, 2021-08 to 2026-01: mean absolute distance in cents between the price the crowd paid and the rate those trades actually won. There are NO regime bands on this chart and that is deliberate - there is no regime state for a calibration curve, and shading it with a market state fitted to something else would assert a relationship this pane has not measured.

Point 1 carries no regime bands here, for two independent reasons and neither is an omission. FIRST, a calibration curve has no regime state: shading it with a market state fitted to something else would assert a relationship this pane has not measured. SECOND, it could not be shaded even if it had one - this leaf is MONTHLY, so five years of it is 60 points at most and this build holds 51, against charts.MIN_POINTS['dateline'] of 60. Below that floor the line is drawn by stepline, which takes no shading at all. The line above is the pane's own leaf over 51 months, 2021-08 to 2026-01, and it is real.

No analog table: no linked price for a calibration curve: a calibration error in cents is not linked to any traded price - it is a property of the crowd across thousands of unrelated markets. Point 2 of the History contract asks for the forward move of the price this leaf is linked to, and this leaf is not linked to one; no other series' episodes are shown in its place.

What changed since last week

leaf	then	now	change	compared with
brier trade time	0.17	0.17	unchanged	2026-10-11
ece	0.01	0.01	unchanged	2026-10-11
log loss	0.50	0.50	unchanged	2026-10-11
longshot volume share latest	0.31	0.31	unchanged	2026-10-11
n resolved markets	3,722,314.00	3,722,314.00	unchanged	2026-10-11
n trades analyzed	121,671,888.00	121,671,888.00	unchanged	2026-10-11

0 of 6 headline leaves moved. A repeated value can mean the upstream had not published, not that nothing happened.

Price analogs are the part of this contract that is missing here, and the reason is printed above rather than left blank.

calibration mad by month · 51 rows

month	mad cents	n
2021-08	1.45	2,282
2021-10	0.52	3,034
2021-11	0.39	2,012
2021-12	0.06	2,462
2022-01	0.90	5,906
2022-04	2.93	4,280
2022-05	1.99	127,214
2022-06	1.96	224,796
2022-07	2.45	229,908
2022-08	1.12	297,818
2022-09	2.02	249,188
2022-10	1.80	208,634
2022-11	2.46	168,694
2022-12	2.06	124,122
2023-01	2.73	53,088
2023-02	2.88	100,578
2023-03	1.23	351,430
2023-04	1.71	271,242
2023-05	2.52	463,642
2023-06	1.35	421,682
2023-07	2.79	443,192
2023-08	2.22	244,042
2023-09	1.59	207,136
2023-10	1.52	316,192
2023-11	2.01	191,146
2023-12	2.43	218,992
2024-01	3.01	385,032
2024-02	1.63	372,168
2024-03	1.88	353,356
2024-04	1.63	337,132
2024-05	1.31	241,464
2024-06	2.11	232,672
2024-07	1.98	249,528
2024-08	1.61	220,996
2024-09	3.67	217,104
2024-10	12.41	870,024
2024-11	5.73	5,123,450
2024-12	0.92	2,016,926
2025-01	0.94	2,427,836
2025-02	2.35	3,364,986
2025-03	4.12	5,529,812
2025-04	3.44	3,930,924
2025-05	2.45	4,726,954
2025-06	2.94	5,838,492
2025-07	1.85	5,694,832
2025-08	1.59	5,908,696
2025-09	2.58	13,549,064
2025-10	2.70	19,697,290
2025-11	1.12	15,550,418
2025-12	1.95	54,292,896
2026-01	1.05	86,603,532

scalars

field	value
n trades analyzed	121,671,888
brier trade time	0.17
ece	0.01
dataset	/

The regime engine Grade D · descriptiveas of 2026-10-19

ReadHistoryData

current state

calm / risk-on

since 2026-04-17

training days

4,296

cross-check

0.064

Markov P(high-vol)

A sparse statistical jump model: k-means clustering with an explicit penalty on switching states, so the label cannot flap day to day the way a plain hidden Markov model's does. Every day is labelled by online inference — using only data available up to that day. We do not publish an in-sample regime chart, because a curve fitted with hindsight describes the past rather than forecasting anything.

selected feature	weight
ret_60	0.777
sortino_60	0.71
dd_20	0.662
iv_vix	0.514
stress_volatility	0.469

The model chooses its own inputs from a wider candidate set; those above are what survived. It is cross-checked against PELT change-point detection and a Markov-switching variance model — and that cross-check has already earned its keep by catching a labelling error of ours before publication.

The jump model labels 4,296 daily observations from 2009-09-18, but regime_model_cache.json caches the state statistics and not the daily labelled path, so this pane's own bands cannot be drawn. The bands below are the shared Markov filter's on the S&P 500. Its own published change-points: 2021-11-22, 2023-03-24, 2025-03-03, 2025-04-29.

Five years of S&P 500 - the series this regime state is computed on - with the high-volatility weeks shaded by the same two-state Markov filter the Regime tab draws. One state, one shading, every tab: the bands are that shared filter's, not a second state fitted for this pane. The filter puts 14 stretches of these five years in the high-volatility state, and those are the shaded ones.

The last 8 times this looked like today — and what S&P 500 did next

episode	z	level	S&P 500 +5d	S&P 500 +21d
2026-08-28	1.63	7,711.76	+0.09%	-0.36%
2026-08-07	1.86	7,757.64	+0.36%	-1.08%
2026-07-17	1.53	7,457.69	-0.61%	+3.85%
2026-06-26	1.51	7,354.02	+2.49%	+0.80%
2026-06-05	1.72	7,383.74	+0.65%	+2.08%
2026-04-29	1.65	7,135.95	+3.21%	+5.99%
2026-04-08	1.06	6,782.81	+3.54%	+8.17%
2026-02-23	1.33	6,837.75	+0.64%	-4.12%

z in [1.0, 2.0) · n=99 episodes · 21-session hit rate 73% · median +5d 0.50, +21d 1.83 · grade C

Bucketed on the S&P 500's own z, not on the jump model's state. Episodes when this series last sat in the same trailing-365-observation z bucket, with the forward 5- and 21-session move of S&P 500. 99 episodes at least 21 days apart; the table shows the most recent 8. A conditional distribution with its count, not a signal.

What changed since last week

leaf	then	now	change	compared with
current state	0.00	0.00	unchanged	2026-10-11
jump penalty	50.00	50.00	unchanged	2026-10-11
median run days	110.00	111.00	+1	2026-10-11
run length days	122.00	127.00	+5	2026-10-11

2 of 4 headline leaves moved. A repeated value can mean the upstream had not published, not that nothing happened.

selected features

key	value
ret_60	0.78
sortino_60	0.71
dd_20	0.66
iv_vix	0.51
stress_volatility	0.47

sample

key	value
start	2009-09-18
end	2026-10-16
rows	4,296
features	17

scalars

field	value
current label	calm / risk-on
training days	4,296
markov p high vol	0.06

How we score ourselves Grade D · descriptiveas of 2026-10-19

ReadHistoryData

our Brier

0.275

base-rate Brier

0.290

the bar to beat

Every forecast carries a hard, machine-checkable threshold and a deadline. At the deadline an official series decides it — there is no judgement call and no quiet deletion. Signals that make a mean-reversion or persistence claim are snapshotted and scored 28 days later on the same terms (398 states recorded, 39 scored so far).

A 28-day score means nothing if the model behind the claim can be refitted, inside those 28 days, on the window it is about to be scored against. So every scoreable claim now takes a sealed holdout over its own forward window: the seal carries a hash of the recorded claim, so the claim cannot be reworded afterwards, and any read of that window before the horizon raises and is written to an audit table whether or not the caller catches it. 0 taken, 0 opened on schedule, 0 refused reads on the record. Across the tracked signals as a family, Hansen's test for superior predictive ability asks whether the best of them beats a coin at all once you account for having looked at every one: not yet computable, because it has nothing to run on.

The benchmark is the unconditional base rate, which is the comparison that can actually embarrass us: beating a coin flip is not a result, beating the base rate is. The ledger →

47 open, 81 resolved. Our Brier 0.275 against a base-rate 0.290.

No five-year line: the leaf here is OUR OWN scoring record and it starts when we started keeping it. 81 forecast(s) resolved and 39 signal state(s) scored so far, from a record beginning 2026-09-04 — a Brier score needs resolutions before it is a number at all, let alone a series. This is the one pane on the site where a short history is the honest headline rather than a caveat.

No analog table: no analog episodes yet, and this is a clock rather than a refusal: 81 resolved bet(s) against a pre-registered floor of 50 (grade_t_thresholds.json). Below that floor this page may not call anything a track record, and printing eight episodes and a hit rate out of what has resolved would be exactly the claim the floor exists to prevent

What changed since last week

leaf	then	now	change	compared with
open › 0 › base rate	0.67	0.67	unchanged	2026-10-11
open › 1 › base rate	0.67	0.67	unchanged	2026-10-11
open › 2 › base rate	0.29	0.29	unchanged	2026-10-11
open › 3 › base rate	0.56	0.24	-0.3197	2026-10-11
open › 4 › base rate	0.49	0.23	-0.257	2026-10-11
recently resolved › 0 › base rate	0.49	0.73	+0.234	2026-10-11
recently resolved › 1 › base rate	0.39	0.66	+0.2717	2026-10-11
recently resolved › 2 › base rate	0.72	0.79	+0.0761	2026-10-11
recently resolved › 3 › base rate	0.72	0.62	-0.1042	2026-10-11
recently resolved › 4 › base rate	0.74	0.49	-0.2498	2026-10-11

7 of 10 headline leaves moved. A repeated value can mean the upstream had not published, not that nothing happened.

Price analogs are the part of this contract that is missing here, and the reason is printed above rather than left blank.

scalars

field	value
open	47
resolved	81
brier	0.27
base rate brier	0.29
signal states	398
signal scored	39

The lockbox — how we avoid fooling ourselves Grade D · descriptiveas of 2026-10-19

ReadData

Our research loop lets an optimiser propose signal rules, but the code that grades them is immutable and the most recent years of data are held back where the optimiser cannot see them. A rule is only kept if it improves on data it was never shown.

held-out excess Sharpe

-5.12

negative = correctly rejected

deflated Sharpe (DSR)

0.00

P(true Sharpe > best-of-1 expectation)

overfit probability (PBO)

not computed

CSCV over the trial ledger; needs 8 trials

SPA p-value (Hansen)

not computed

the ledger holds one trial; a family of one has no maximum to test

lockbox window

2024-01-01 → 2025-03-05

never shown to the optimiser

We publish this because a research process that has never rejected anything is not a process. The figure above is a rejection: the rule looked good on the data it was fitted to and did not survive contact with the years it had never seen.

scalars

field	value
lockbox excess sharpe	-5.12
lockbox window	2024-01-01..2025-03-05
lockbox dsr	0.00
lockbox sr0 annualised	0.00
n trials	1
pbo	not reported

What actually happens after a signal fires Grade C · counted distributionas of 2026-10-18 11:56 UTC

ReadHistoryData

Cumulative abnormal return after each trigger, measured against the sample's own drift, with 95% confidence intervals from 2,000 bootstrap resamples. Only results whose interval excludes zero are listed — and the sample size is shown beside each, because a confidence interval on seven events is a hint while one on fifty is a finding.

trigger	market	horizon	mean CAR	95% CI	t	hit rate	n	confidence
Implied vol below realised (the premium inverts)	S&P 500 since 2018-02-14	+1d	-0.71%	[-1.4, -0.2]	-2.27	36%	53	reasonable
Managed money at a 3y positioning extreme (long)	Silver since 2009-12-04	+5d	-2.42%	[-3.9, -1.0]	-3.11	27%	22	thin
Managed money at a 3y positioning extreme (long)	Silver since 2009-12-04	+10d	-3.01%	[-5.1, -0.7]	-2.64	27%	22	thin
Managed money at a 3y positioning extreme (long)	Silver since 2009-12-04	+20d	-5.74%	[-10.1, -1.9]	-2.61	23%	22	thin
Managed money at a 3y positioning extreme (short)	EUR/USD since 2010-07-20	+20d	-1.22%	[-2.5, -0.1]	-2.0	32%	19	thin
Managed money at a 3y positioning extreme (long)	JPY since 2010-07-20	+10d	-1.06%	[-2.0, -0.2]	-2.34	25%	16	thin

Of 64 relationships tested, 6 clear the significance bar and only those with n≥30 are worth leaning on. The 2s10s inversion — the most widely repeated of these relationships — shows no significant effect at any horizon we measured, which is the sort of result that only appears if you look.

64 relationship(s) tested, 6 clearing the interval, as of 2026-10-18 11:56 UTC.

No five-year line: this panel's leaf is a cumulative abnormal return per TRIGGER, not a reading that exists on every date. 64 relationship(s) were tested and each publishes one number per horizon; there is no series joining them, and the events themselves are decades apart in places (the 2s10s inversion sample reaches back to 1978).

Implied vol below realised (the premium inverts) — S&P 500, and what ^GSPC did next

event	+1d	+5d	+10d	+20d
2026-06-30	-0.24%	-0.34%	+0.72%	-2.96%
2026-06-18	-0.40%	-2.10%	+0.25%	-1.26%
2026-04-09	-0.14%	+3.00%	+3.83%	+6.75%
2025-04-09	-3.55%	-3.50%	+0.26%	+3.23%
2025-03-19	-0.24%	+0.53%	-0.32%	-7.79%
2025-01-15	-0.24%	+2.68%	+1.77%	+2.25%
2024-11-29	+0.22%	+0.83%	+0.06%	-2.59%
2024-08-09	-0.02%	+3.73%	+5.05%	+1.86%

The 8 most recent of n=53 usable event(s) · hit rate at +20d 58.3%, over the 24 event(s) whose +20d window has closed and whose return is defined, not all 53 — not the 8 printed here · cumulative abnormal return, percent, re-based to 0 on the event day · grade C

Managed money at a 3y positioning extreme (long) — Silver, and what SI=F did next

event	+1d	+5d	+10d	+20d
2025-06-06	+1.79%	+0.52%	-0.02%	+0.88%
2025-03-14	-0.37%	-2.85%	+0.95%	-7.91%
2024-10-25	+0.64%	-3.40%	-7.29%	-7.82%
2024-09-20	-1.36%	+0.91%	+2.63%	+5.03%
2024-07-12	-0.73%	-6.18%	-10.69%	-12.41%
2024-05-31	+1.09%	-3.41%	-3.37%	-4.09%
2024-05-17	+3.62%	-2.52%	-1.69%	-6.45%
2024-04-12	+1.35%	+1.75%	-4.03%	-0.68%

The 8 most recent of n=22 usable event(s) · hit rate at +20d 22.7%, over the 22 event(s) whose +20d window has closed and whose return is defined — not the 8 printed here · cumulative abnormal return, percent, re-based to 0 on the event day · grade C

Managed money at a 3y positioning extreme (short) — EUR/USD, and what EURUSD=X did next

event	+1d	+5d	+10d	+20d
2026-06-26	+0.21%	+0.54%	+0.64%	+0.15%
2026-03-27	-0.37%	+0.06%	+1.36%	+1.30%
2025-01-31	-1.45%	-0.11%	+0.63%	+0.00%
2024-11-01	-0.11%	-0.83%	-3.31%	-2.99%
2024-04-19	+0.13%	+0.80%	+0.84%	+2.10%
2020-01-31	+0.55%	-0.41%	-1.66%	-0.27%
2019-12-30	+0.18%	-0.18%	-0.58%	-1.39%
2019-04-26	+0.12%	+0.33%	+0.77%	+0.67%

The 8 most recent of n=19 usable event(s) · hit rate at +20d 26.3%, over the 19 event(s) whose +20d window has closed and whose return is defined — not the 8 printed here · cumulative abnormal return, percent, re-based to 0 on the event day · grade C

Managed money at a 3y positioning extreme (long) — JPY, and what JPY=X did next

event	+1d	+5d	+10d	+20d
2026-09-18	+0.58%	+1.68%	+1.10%	+1.10%
2025-03-28	-0.82%	-3.13%	-4.86%	-6.00%
2025-02-28	+0.50%	-1.30%	-1.43%	+0.47%
2024-08-09	-0.43%	+1.03%	-1.04%	-3.02%
2020-11-16	-0.13%	-0.84%	-0.61%	-0.79%
2019-08-16	+0.32%	+0.34%	+0.29%	+1.87%
2019-08-02	-0.80%	-1.48%	-1.30%	-1.01%
2016-01-08	-0.40%	+0.54%	+0.28%	-0.66%

The 8 most recent of n=16 usable event(s) · hit rate at +20d 43.8%, over the 16 event(s) whose +20d window has closed and whose return is defined — not the 8 printed here · cumulative abnormal return, percent, re-based to 0 on the event day · grade C

4 of the 16 relationship(s) carrying dated events are shown: the ones whose sample-wide result cleared the confidence interval, which is the same cut the Read tab makes. The other 12 are on the Read tab with their counts and intervals. Each table is that relationship's OWN events, not a pooled sample — the triggers are different questions and pooling them would produce a hit rate for no relationship in particular.

What changed since last week

leaf	then	now	change	compared with
window › after	20.00	20.00	unchanged	2026-10-11
window › before	10.00	10.00	unchanged	2026-10-11

0 of 2 headline leaves moved. A repeated value can mean the upstream had not published, not that nothing happened.

results · 64 rows

event	subject	n events	first event	horizon days	mean car pct	median car pct	ci95 pct	t stat	hit rate
Managed money at a 3y positioning extreme (long)	Crude (WTI)	15	2011-03-18	1	not reported	not reported	not reported	not reported	0.47
Managed money at a 3y positioning extreme (long)	Crude (WTI)	15	2011-03-18	5	not reported	not reported	not reported	not reported	0.40
Managed money at a 3y positioning extreme (long)	Crude (WTI)	15	2011-03-18	10	not reported	not reported	not reported	not reported	0.40
Managed money at a 3y positioning extreme (long)	Crude (WTI)	15	2011-03-18	20	not reported	not reported	not reported	not reported	0.47
Managed money at a 3y positioning extreme (long)	Nat gas (HH)	7	2009-12-04	1	0.44	-0.80	-1.24, 3.07	0.34	0.29
Managed money at a 3y positioning extreme (long)	Nat gas (HH)	7	2009-12-04	5	0.21	-2.31	-3.65, 4.78	0.09	0.43
Managed money at a 3y positioning extreme (long)	Nat gas (HH)	7	2009-12-04	10	1.97	-3.23	-4.21, 10.02	0.49	0.43
Managed money at a 3y positioning extreme (long)	Nat gas (HH)	7	2009-12-04	20	-0.56	-1.37	-8.10, 7.77	-0.13	0.43
Managed money at a 3y positioning extreme (long)	Gold	23	2009-12-04	1	0.12	0.04	-0.23, 0.46	0.65	0.52
Managed money at a 3y positioning extreme (long)	Gold	23	2009-12-04	5	-0.55	-0.54	-1.69, 0.57	-0.93	0.43
Managed money at a 3y positioning extreme (long)	Gold	23	2009-12-04	10	0.32	1.41	-0.95, 1.54	0.48	0.65
Managed money at a 3y positioning extreme (long)	Gold	23	2009-12-04	20	0.78	-0.39	-1.59, 3.11	0.64	0.48
Managed money at a 3y positioning extreme (long)	Silver	22	2009-12-04	1	0.01	-0.12	-0.63, 0.66	0.02	0.46
Managed money at a 3y positioning extreme (long)	Silver	22	2009-12-04	5	-2.42	-2.44	-3.92, -0.99	-3.11	0.27
Managed money at a 3y positioning extreme (long)	Silver	22	2009-12-04	10	-3.01	-3.57	-5.06, -0.69	-2.64	0.27
Managed money at a 3y positioning extreme (long)	Silver	22	2009-12-04	20	-5.74	-5.41	-10.09, -1.93	-2.61	0.23
Managed money at a 3y positioning extreme (long)	Copper	24	2009-12-04	1	0.06	0.08	-0.33, 0.52	0.29	0.54
Managed money at a 3y positioning extreme (long)	Copper	24	2009-12-04	5	0.03	-0.25	-1.32, 1.46	0.04	0.46
Managed money at a 3y positioning extreme (long)	Copper	24	2009-12-04	10	0.34	-0.19	-1.41, 2.16	0.37	0.42
Managed money at a 3y positioning extreme (long)	Copper	24	2009-12-04	20	0.24	1.49	-2.48, 3.01	0.17	0.54
Managed money at a 3y positioning extreme (long)	EUR/USD	23	2010-07-20	1	-0.07	-0.08	-0.23, 0.08	-0.84	0.43
Managed money at a 3y positioning extreme (long)	EUR/USD	23	2010-07-20	5	0.08	0.01	-0.28, 0.44	0.43	0.52
Managed money at a 3y positioning extreme (long)	EUR/USD	23	2010-07-20	10	0.02	-0.01	-0.54, 0.66	0.06	0.48
Managed money at a 3y positioning extreme (long)	EUR/USD	23	2010-07-20	20	0.30	-0.15	-0.39, 1.07	0.78	0.48
Managed money at a 3y positioning extreme (long)	JPY	16	2010-07-20	1	-0.15	-0.08	-0.41, 0.10	-1.11	0.44
Managed money at a 3y positioning extreme (long)	JPY	16	2010-07-20	5	-0.22	-0.02	-0.98, 0.51	-0.56	0.50
Managed money at a 3y positioning extreme (long)	JPY	16	2010-07-20	10	-1.06	-0.83	-2.00, -0.22	-2.34	0.25
Managed money at a 3y positioning extreme (long)	JPY	16	2010-07-20	20	-0.83	-0.66	-1.86, 0.14	-1.56	0.44
Managed money at a 3y positioning extreme (short)	Crude (WTI)	24	2010-09-24	1	-0.05	0.05	-0.82, 0.70	-0.13	0.54
Managed money at a 3y positioning extreme (short)	Crude (WTI)	24	2010-09-24	5	0.22	0.69	-1.83, 2.18	0.21	0.54
Managed money at a 3y positioning extreme (short)	Crude (WTI)	24	2010-09-24	10	-0.09	0.30	-2.46, 2.27	-0.07	0.54
Managed money at a 3y positioning extreme (short)	Crude (WTI)	24	2010-09-24	20	0.74	1.80	-2.32, 3.50	0.50	0.58
Managed money at a 3y positioning extreme (short)	Nat gas (HH)	25	2010-10-15	1	0.37	0.47	-1.44, 2.18	0.39	0.56
Managed money at a 3y positioning extreme (short)	Nat gas (HH)	25	2010-10-15	5	0.74	-0.15	-2.83, 3.60	0.44	0.48
Managed money at a 3y positioning extreme (short)	Nat gas (HH)	25	2010-10-15	10	-0.70	0.86	-6.77, 4.79	-0.24	0.52
Managed money at a 3y positioning extreme (short)	Nat gas (HH)	25	2010-10-15	20	-3.66	-1.83	-11.44, 2.84	-0.96	0.44
Managed money at a 3y positioning extreme (short)	Gold	22	2011-12-23	1	0.26	0.29	-0.17, 0.68	1.18	0.59
Managed money at a 3y positioning extreme (short)	Gold	22	2011-12-23	5	0.83	0.31	-0.03, 1.73	1.83	0.64
Managed money at a 3y positioning extreme (short)	Gold	22	2011-12-23	10	0.20	-0.60	-1.14, 1.62	0.28	0.46
Managed money at a 3y positioning extreme (short)	Gold	22	2011-12-23	20	-1.24	-1.86	-2.96, 0.56	-1.33	0.27
Managed money at a 3y positioning extreme (short)	Silver	28	2009-12-04	1	0.24	-0.02	-0.38, 0.88	0.74	0.50
Managed money at a 3y positioning extreme (short)	Silver	28	2009-12-04	5	0.67	0.06	-0.95, 2.47	0.77	0.50
Managed money at a 3y positioning extreme (short)	Silver	28	2009-12-04	10	1.07	0.05	-1.01, 3.21	0.97	0.50
Managed money at a 3y positioning extreme (short)	Silver	28	2009-12-04	20	0.78	0.63	-1.85, 3.33	0.58	0.54
Managed money at a 3y positioning extreme (short)	Copper	27	2009-12-04	1	-0.30	-0.18	-0.68, 0.06	-1.58	0.37
Managed money at a 3y positioning extreme (short)	Copper	27	2009-12-04	5	-0.33	-0.43	-1.25, 0.62	-0.69	0.44
Managed money at a 3y positioning extreme (short)	Copper	27	2009-12-04	10	-0.34	-0.09	-1.54, 0.79	-0.56	0.48
Managed money at a 3y positioning extreme (short)	Copper	27	2009-12-04	20	-0.04	0.53	-1.50, 1.35	-0.05	0.52
Managed money at a 3y positioning extreme (short)	EUR/USD	19	2010-07-20	1	-0.17	-0.09	-0.39, 0.05	-1.46	0.37
Managed money at a 3y positioning extreme (short)	EUR/USD	19	2010-07-20	5	-0.09	-0.11	-0.54, 0.39	-0.39	0.47
Managed money at a 3y positioning extreme (short)	EUR/USD	19	2010-07-20	10	-0.28	0.09	-1.18, 0.59	-0.62	0.53
Managed money at a 3y positioning extreme (short)	EUR/USD	19	2010-07-20	20	-1.22	-0.68	-2.47, -0.12	-2.00	0.32
Managed money at a 3y positioning extreme (short)	JPY	28	2010-07-20	1	0.10	0.08	-0.10, 0.32	0.97	0.57
Managed money at a 3y positioning extreme (short)	JPY	28	2010-07-20	5	0.21	0.15	-0.21, 0.67	0.97	0.50
Managed money at a 3y positioning extreme (short)	JPY	28	2010-07-20	10	-0.14	0.16	-0.82, 0.55	-0.41	0.54
Managed money at a 3y positioning extreme (short)	JPY	28	2010-07-20	20	0.00	0.02	-1.13, 1.29	0.00	0.50
2s10s crosses into inversion	S&P 500	44	1978-08-18	1	0.03	0.03	-0.18, 0.27	0.30	0.55
2s10s crosses into inversion	S&P 500	44	1978-08-18	5	-0.13	-0.25	-0.68, 0.40	-0.48	0.46
2s10s crosses into inversion	S&P 500	44	1978-08-18	10	0.03	0.41	-0.77, 0.78	0.07	0.59
2s10s crosses into inversion	S&P 500	44	1978-08-18	20	0.37	1.04	-0.72, 1.42	0.66	0.64
Implied vol below realised (the premium inverts)	S&P 500	53	2018-02-14	1	-0.71	-0.18	-1.45, -0.17	-2.27	0.36
Implied vol below realised (the premium inverts)	S&P 500	53	2018-02-14	5	-0.68	0.53	-1.93, 0.38	-1.18	0.57
Implied vol below realised (the premium inverts)	S&P 500	53	2018-02-14	10	-0.23	0.81	-1.75, 1.01	-0.32	0.72
Implied vol below realised (the premium inverts)	S&P 500	53	2018-02-14	20	0.10	1.95	-1.54, 1.59	0.12	0.68

scalars

field	value
n relationships	64
n significant	6

The reading pipeline Grade D · descriptiveas of 2026-10-19

ReadData

articles ingested

5,263

discovered

5,838

the rest queue for later runs

Independent macro and finance writers are fetched twice daily, parsed to text and embedded so the commentary can cite what they actually argued rather than paraphrasing a headline. Where they disagree with each other, the weekly note names both sides. This week's →

scalars

field	value
articles ingested	5,263
discovered	5,838

Through time — level, distribution, seasonality and range Grade C · counted distributionas of 2026-10-19 11:47 UTC

ReadHistoryResidual checkConditionalData

36 headline series across 10 desks, each read four ways: level against its own trailing mean, position in its full history, calendar-month seasonality, and an empirical 63-day range. The most stretched is Copper at the 100th percentile of 415 observations back to 1992-01-01. Only 3 of 36 show a calendar pattern larger than their own noise (Gasoline (Gulf), 2s10s slope, 5y breakeven) — the rest are shown with that stated, because a seasonal chart drawn on noise is the most convincing bad chart in finance. Behind the tabs: 18 distinct series decomposed with statsmodels 0.14.6, 18 with a two-state Markov filter and 18 with an analog episode table scored against a linked price.

10y TreasuryWTI crudeBroad dollarVIX

level

5.10

pct · as of 2026-10-15

percentile

of 16,182 obs since 1962

percentile, 5y

of 1,248 obs

vs 13w mean

+0.19

level minus its own quarter

vs 52w mean

+0.67

level minus its own year

year on year

+1.08

same week last year

TrendDistributionSeasonalityRangeRegime

MSTL with periods 5 and 52 on 3381 weekly observations from 1962-01-05. The trend is drawn over the level and the remainder under it; nothing here is extrapolated. The trend explains 99% of the variance the remainder does not (Hyndman's trend strength); the seasonal component explains 14%.

What the trend and the seasonal do not account for, on the same dates. A remainder that trends is a decomposition that has missed something, which is why it is drawn rather than described.

Harvey–Jaeger local linear trend on the quarterly average level, 1986-12-31 to 2026-09-30 (160 quarters), smoothed: each quarter's estimate uses the quarters after it.

No cycle is drawn: the implied period is 1.5 years, pinned at the 1.5–12-year bound, so the data do not locate a cycle inside it; the cycle does not earn its parameters: BIC 168.1 with it against 164.8 for the trend alone; the two optimisers put the period at 1.5 and 4.9 years, so it is not identified (best period 1.5 y, damping 0.49). The trend above is the trend-only model, which the data prefer. The trend-only model assigns no variance to the irregular, so its trend coincides with the quarterly level: at this cadence the model finds nothing to separate from the path itself.

The whole history as a distribution, with today marked. Median 5.39 over 16,182 observations.

The same record read cumulatively: today sits at the 47 percentile of the full history (n=16,182) and the 99 of the last five years (n=1,248). A level extreme against forty years and ordinary against five is a different fact from one that is extreme against both.

The spread of monthly means is no larger than the noise — this series has no seasonal worth trading. Kruskal-Wallis across calendar months; the corner figure is the average number of years behind each bar.

The MSTL seasonal component itself, at periods 5 and 52. Drawn rather than summarised because a seasonal that changes shape year to year is not one a reader can plan against, and the chart shows that where a single average would hide it.

Seasonal diagnostic: The STL seasonal component's own mean at each calendar month, across 65 years, against the year-to-year standard deviation at that same month (ratio 0.44). The year-to-year noise at these calendar months is as large as the mean seasonal effect itself, so the shape should be read as unstable, not as a repeating pattern.

month	market holidays	Crude prints	Gas prints	option expiries	quad-witchings
Jan	19	45	43	10	0
Feb	10	40	41	10	0
Mar	3	45	45	10	10
Apr	7	42	42	10	0
May	10	43	43	10	0
Jun	4	45	45	10	10
Jul	10	44	43	10	0
Aug	0	45	45	10	0
Sep	10	42	43	10	10
Oct	0	44	44	10	0
Nov	10	44	43	10	0
Dec	10	43	44	10	10

Event days per calendar month, 2016–2025. Every column is derived from a published rule, so it was knowable years ahead: the whole table is stamped available_at 1900-01-01, never the build date. Dropped, because they are announced rather than derived and the schedule loaded here does not cover the window: FOMC decisions, CPI prints.

The high-minus-low width of the trailing 88-observation window (the same 63-day horizon as the band above), against the last 1,248 such windows over five years. Today's window sits at the 71 percentile of its own five-year range history; the 10th-to-90th of that history is 0.45 to 1.37 (level). A description of how wide the market has recently traded, not a target.

Where this has historically gone over 63 trading days: 10th to 90th percentile of 16,094 realised moves, applied to today. A base rate for sizing, not a forecast.

downside (10th)

4.30

-0.80

median path

5.13

50th percentile

upside (90th)

5.84

+0.74

Markov-filtered mean

+0.001

per week, state-weighted

The band above is unconditional: it does not know what state we are in. The filtered mean beside it does — it is the two-state Markov model's mean weighted by today's filtered probability (15 of the high-volatility state). The two disagree when the current state is unusual, which is the point of printing both.

state

low volatility

since 2025-07-04

P(high vol) now

filtered, this week

persisted

weeks (476 days)

P(exit in 4w)

0.0%

0 of 3 spells this old

spells on record

24 completed in this state

Two-state Markov-switching model with switching variance, fitted on 1300 weekly change in level. The line is the FILTERED probability of the high-volatility state — a statement about weeks that have happened, not a forecast. Exit counts are empirical spell counts, not the model's expected durations.

The same weeks on the level: shaded stretches are the ones the filter puts in the high-volatility state. Mean +0.002 per week at 0.086 standard deviation in the calm state against -0.001 at 0.147 in the volatile one.

Markov-switching AR(1) (MarkovAutoregression(k_regimes=2, order=1, switching_ar=False, switching_variance=True), 779 periods since 2011-11-18): low volatility now, filtered p(high) 0.29; AR(1) -0.092; sd 0.145 high against 0.084 low. The two Markov filters put the same state on 96% of 260 common periods; the MarkovRegression reads low volatility now. The filtered probability is about periods that have happened.

level

101.91

usd · as of 2026-10-14

percentile

of 9,523 obs since 1986

percentile, 5y

of 1,247 obs

vs 13w mean

+10.17

level minus its own quarter

vs 52w mean

+21.20

level minus its own year

year on year

+74.8%

same week last year

TrendDistributionSeasonalityRangeRegime

MSTL with periods 5 and 52 on 2126 weekly observations from 1986-01-03. The trend is drawn over the level and the remainder under it; nothing here is extrapolated. The trend explains 96% of the variance the remainder does not (Hyndman's trend strength); the seasonal component explains 16%.

What the trend and the seasonal do not account for, on the same dates. A remainder that trends is a decomposition that has missed something, which is why it is drawn rather than described.

Harvey–Jaeger local linear trend on the quarterly average log level, 1986-12-31 to 2026-09-30 (160 quarters), smoothed: each quarter's estimate uses the quarters after it.

No cycle is drawn: the cycle does not earn its parameters: BIC -109.0 with it against -126.7 for the trend alone (best period 3.2 y, damping 0.81). The trend above is the trend-only model, which the data prefer. The trend-only model assigns no variance to the irregular, so its trend coincides with the quarterly level: at this cadence the model finds nothing to separate from the path itself.

The whole history as a distribution, with today marked. Median 47.11 over 9,523 observations.

The same record read cumulatively: today sits at the 95 percentile of the full history (n=9,523) and the 92 of the last five years (n=1,247). A level extreme against forty years and ordinary against five is a different fact from one that is extreme against both.

Seasonal diagnostic: The STL seasonal component's own mean at each calendar month, across 41 years, against the year-to-year standard deviation at that same month (ratio 0.71). The year-to-year noise at these calendar months is as large as the mean seasonal effect itself, so the shape should be read as unstable, not as a repeating pattern.

month	market holidays	Crude prints	Gas prints	option expiries	quad-witchings
Jan	19	45	43	10	0
Feb	10	40	41	10	0
Mar	3	45	45	10	10
Apr	7	42	42	10	0
May	10	43	43	10	0
Jun	4	45	45	10	10
Jul	10	44	43	10	0
Aug	0	45	45	10	0
Sep	10	42	43	10	10
Oct	0	44	44	10	0
Nov	10	44	43	10	0
Dec	10	43	44	10	10

The high-minus-low width of the trailing 88-observation window (the same 63-day horizon as the band above), against the last 1,247 such windows over five years. Today's window sits at the 83 percentile of its own five-year range history; the 10th-to-90th of that history is 19.95 to 64.63 (% of level). A description of how wide the market has recently traded, not a target.

Where this has historically gone over 63 trading days: 10th to 90th percentile of 9,435 realised moves, applied to today. A base rate for sizing, not a forecast.

downside (10th)

80.66

-20.9%

median path

104.75

50th percentile

upside (90th)

133.05

+30.6%

Markov-filtered mean

-0.419

per week, state-weighted

The band above is unconditional: it does not know what state we are in. The filtered mean beside it does — it is the two-state Markov model's mean weighted by today's filtered probability (38 of the high-volatility state). The two disagree when the current state is unusual, which is the point of printing both.

state

low volatility

since 2026-10-16

P(high vol) now

filtered, this week

persisted

weeks (7 days)

P(exit in 4w)

35.3%

6 of 17 spells this old

spells on record

17 completed in this state

Two-state Markov-switching model with switching variance, fitted on 1300 weekly % log change. The line is the FILTERED probability of the high-volatility state — a statement about weeks that have happened, not a forecast. Exit counts are empirical spell counts, not the model's expected durations.

The same weeks on the level: shaded stretches are the ones the filter puts in the high-volatility state. Mean +0.274 per week at 4.304 standard deviation in the calm state against -1.569 at 15.303 in the volatile one.

Markov-switching AR(1) (MarkovAutoregression(k_regimes=2, order=1, switching_ar=False, switching_variance=True), 779 periods since 2011-11-18): low volatility now, filtered p(high) 0.43; AR(1) +0.004; sd 14.405 high against 4.119 low. The two Markov filters put the same state on 97% of 260 common periods; the MarkovRegression reads low volatility now. The filtered probability is about periods that have happened.

level

121.60

idx · as of 2026-10-09

percentile

of 5,208 obs since 2006

percentile, 5y

of 1,250 obs

vs 13w mean

+2.00

level minus its own quarter

vs 52w mean

+1.88

level minus its own year

year on year

+0.4%

same week last year

TrendDistributionSeasonalityRangeRegime

MSTL with periods 5 and 52 on 1084 weekly observations from 2006-01-06. The trend is drawn over the level and the remainder under it; nothing here is extrapolated. The trend explains 98% of the variance the remainder does not (Hyndman's trend strength); the seasonal component explains 11%.

What the trend and the seasonal do not account for, on the same dates. A remainder that trends is a decomposition that has missed something, which is why it is drawn rather than described.

Harvey–Jaeger local linear trend on the quarterly average log level, 2006-03-31 to 2026-09-30 (83 quarters), smoothed: each quarter's estimate uses the quarters after it.

No cycle is drawn: the cycle does not earn its parameters: BIC -328.5 with it against -348.1 for the trend alone (best period 2.6 y, damping 0.82). The trend above is the trend-only model, which the data prefer. The trend-only model assigns no variance to the irregular, so its trend coincides with the quarterly level: at this cadence the model finds nothing to separate from the path itself.

The whole history as a distribution, with today marked. Median 110.25 over 5,208 observations.

The same record read cumulatively: today sits at the 91 percentile of the full history (n=5,208) and the 65 of the last five years (n=1,250). A level extreme against forty years and ordinary against five is a different fact from one that is extreme against both.

Seasonal diagnostic: The STL seasonal component's own mean at each calendar month, across 21 years, against the year-to-year standard deviation at that same month (ratio 0.91). The year-to-year noise at these calendar months is as large as the mean seasonal effect itself, so the shape should be read as unstable, not as a repeating pattern.

month	market holidays	Crude prints	Gas prints	option expiries	quad-witchings
Jan	19	45	43	10	0
Feb	10	40	41	10	0
Mar	3	45	45	10	10
Apr	7	42	42	10	0
May	10	43	43	10	0
Jun	4	45	45	10	10
Jul	10	44	43	10	0
Aug	0	45	45	10	0
Sep	10	42	43	10	10
Oct	0	44	44	10	0
Nov	10	44	43	10	0
Dec	10	43	44	10	10

The high-minus-low width of the trailing 88-observation window (the same 63-day horizon as the band above), against the last 1,250 such windows over five years. Today's window sits at the 38 percentile of its own five-year range history; the 10th-to-90th of that history is 2.63 to 6.89 (% of level). A description of how wide the market has recently traded, not a target.

Where this has historically gone over 63 trading days: 10th to 90th percentile of 5,120 realised moves, applied to today. A base rate for sizing, not a forecast.

downside (10th)

116.72

-4.0%

median path

121.75

50th percentile

upside (90th)

127.68

+5.0%

Markov-filtered mean

-0.001

per week, state-weighted

The band above is unconditional: it does not know what state we are in. The filtered mean beside it does — it is the two-state Markov model's mean weighted by today's filtered probability (1 of the high-volatility state). The two disagree when the current state is unusual, which is the point of printing both.

state

low volatility

since 2020-11-20

P(high vol) now

filtered, this week

persisted

308

weeks (2156 days)

P(exit in 4w)

not countable

0 of 0 spells this old

spells on record

20 completed in this state

Two-state Markov-switching model with switching variance, fitted on 1083 weekly % log change. The line is the FILTERED probability of the high-volatility state — a statement about weeks that have happened, not a forecast. Exit counts are empirical spell counts, not the model's expected durations.

The same weeks on the level: shaded stretches are the ones the filter puts in the high-volatility state. Mean -0.004 per week at 0.642 standard deviation in the calm state against +0.266 at 1.559 in the volatile one.

Markov-switching AR(1) (MarkovAutoregression(k_regimes=2, order=1, switching_ar=False, switching_variance=True), 779 periods since 2011-11-11): low volatility now, filtered p(high) 0.08; AR(1) +0.045; sd 1.024 high against 0.577 low. The two Markov filters put the same state on 90% of 260 common periods; the MarkovRegression reads low volatility now. The filtered probability is about periods that have happened.

level

16.46

pct · as of 2026-10-15

percentile

of 9,296 obs since 1990

percentile, 5y

of 1,287 obs

vs 13w mean

+1.08

level minus its own quarter

vs 52w mean

-1.43

level minus its own year

year on year

-4.32

same week last year

TrendDistributionSeasonalityRangeRegime

MSTL with periods 5 and 52 on 1920 weekly observations from 1990-01-05. The trend is drawn over the level and the remainder under it; nothing here is extrapolated. The trend explains 67% of the variance the remainder does not (Hyndman's trend strength); the seasonal component explains 16%.

What the trend and the seasonal do not account for, on the same dates. A remainder that trends is a decomposition that has missed something, which is why it is drawn rather than described.

Harvey–Jaeger local linear trend on the quarterly average level, 1990-03-31 to 2026-09-30 (147 quarters), smoothed: each quarter's estimate uses the quarters after it.

No cycle is drawn: the implied period is 1.5 years, pinned at the 1.5–12-year bound, so the data do not locate a cycle inside it (best period 1.5 y, damping 0.59). The trend above is the trend-only model, which the data prefer.

The whole history as a distribution, with today marked. Median 17.56 over 9,296 observations.

The same record read cumulatively: today sits at the 43 percentile of the full history (n=9,296) and the 38 of the last five years (n=1,287). A level extreme against forty years and ordinary against five is a different fact from one that is extreme against both.

Seasonal diagnostic: The STL seasonal component's own mean at each calendar month, across 37 years, against the year-to-year standard deviation at that same month (ratio 0.62). The year-to-year noise at these calendar months is as large as the mean seasonal effect itself, so the shape should be read as unstable, not as a repeating pattern.

month	market holidays	Crude prints	Gas prints	option expiries	quad-witchings
Jan	19	45	43	10	0
Feb	10	40	41	10	0
Mar	3	45	45	10	10
Apr	7	42	42	10	0
May	10	43	43	10	0
Jun	4	45	45	10	10
Jul	10	44	43	10	0
Aug	0	45	45	10	0
Sep	10	42	43	10	10
Oct	0	44	44	10	0
Nov	10	44	43	10	0
Dec	10	43	44	10	10

The high-minus-low width of the trailing 88-observation window (the same 63-day horizon as the band above), against the last 1,287 such windows over five years. Today's window sits at the 3 percentile of its own five-year range history; the 10th-to-90th of that history is 7.37 to 26.63 (level). A description of how wide the market has recently traded, not a target.

Where this has historically gone over 63 trading days: 10th to 90th percentile of 9,208 realised moves, applied to today. A base rate for sizing, not a forecast.

downside (10th)

8.99

-7.47

median path

15.86

50th percentile

upside (90th)

24.10

+7.64

Markov-filtered mean

-0.131

per week, state-weighted

The band above is unconditional: it does not know what state we are in. The filtered mean beside it does — it is the two-state Markov model's mean weighted by today's filtered probability (4 of the high-volatility state). The two disagree when the current state is unusual, which is the point of printing both.

state

low volatility

since 2026-06-19

P(high vol) now

filtered, this week

persisted

weeks (126 days)

P(exit in 4w)

11.1%

2 of 18 spells this old

spells on record

172

86 completed in this state

The same weeks on the level: shaded stretches are the ones the filter puts in the high-volatility state. Mean -0.155 per week at 1.676 standard deviation in the calm state against +0.402 at 5.715 in the volatile one.

Markov-switching AR(1) (MarkovAutoregression(k_regimes=2, order=1, switching_ar=False, switching_variance=True), 779 periods since 2011-11-18): low volatility now, filtered p(high) 0.06; AR(1) -0.119; sd 5.678 high against 1.626 low. The two Markov filters put the same state on 95% of 260 common periods; the MarkovRegression reads low volatility now. The filtered probability is about periods that have happened.

10y TreasuryWTI crudeBroad dollarVIX

10y Treasury · last 5.10 as of 2026-10-15 · 16,182 observations from 1962-01-02

Five years of 10y Treasury with the high-volatility weeks shaded, from the same two-state filter the Regime tab draws. One state, one shading, every tab.

The last 8 times this looked like today — and what S&P 500 did next

episode	z	level	S&P 500 +5d	S&P 500 +21d
2026-09-14	3.13	4.97	+1.90%	+2.43%
2026-08-24	2.15	4.70	+0.43%	+1.46%
2026-08-03	2.43	4.70	+2.01%	+0.41%
2026-07-13	2.09	4.62	-0.96%	+2.83%
2026-05-19	2.23	4.67	+2.25%	+0.90%
2023-10-12	2.30	4.70	-1.65%	+1.51%
2023-09-21	2.11	4.49	-0.70%	-2.44%
2022-11-08	2.16	4.14	+4.27%	+2.76%

z in [2.0, 99.0) · n=141 episodes · 21-session hit rate 10% · median +5d 0.00, +21d 0.00 · grade C

Episodes when this series last sat in the same trailing-365-observation z bucket, with the forward 5- and 21-session move of S&P 500. 141 episodes at least 21 days apart; the table shows the most recent 8. A conditional distribution with its count, not a signal.

WTI crude · last 101.91 as of 2026-10-14 · 9,523 observations from 1986-01-02

Five years of WTI crude with the high-volatility weeks shaded, from the same two-state filter the Regime tab draws. One state, one shading, every tab.

The last 8 times this looked like today — and what WTI crude did next

episode	z	level	WTI crude +5d	WTI crude +21d
2026-09-09	1.60	97.26	+6.54%	+1.03%
2026-08-19	1.02	87.28	-4.38%	+18.25%
2026-07-21	1.01	86.04	-5.96%	+1.44%
2026-05-26	1.99	97.63	-0.16%	-26.85%
2026-04-17	1.72	85.91	+14.56%	+30.66%
2026-03-03	1.58	74.48	+12.39%	+36.82%
2024-07-01	1.13	84.70	-1.26%	-10.07%
2024-04-23	1.04	84.17	-0.81%	-5.96%

z in [1.0, 2.0) · n=224 episodes · 21-session hit rate 56% · median +5d -0.16, +21d 1.41 · grade C

Episodes when this series last sat in the same trailing-365-observation z bucket, with the forward 5- and 21-session move of WTI crude. 224 episodes at least 21 days apart; the table shows the most recent 8. A conditional distribution with its count, not a signal.

Broad dollar · last 121.60 as of 2026-10-09 · 5,208 observations from 2006-01-02

Five years of Broad dollar with the high-volatility weeks shaded, from the same two-state filter the Regime tab draws. One state, one shading, every tab.

The last 8 times this looked like today — and what Broad dollar did next

episode	z	level	Broad dollar +5d	Broad dollar +21d
2025-03-31	1.16	126.71	+0.10%	-3.25%
2025-03-10	1.07	126.27	-0.71%	+0.47%
2025-02-14	1.53	127.15	-0.01%	-1.40%
2025-01-24	1.89	127.46	+0.80%	-0.21%
2024-10-31	1.95	124.80	-0.02%	+1.43%
2024-10-10	1.11	123.10	+0.83%	+2.02%
2024-09-03	1.01	122.78	+0.18%	-0.98%
2024-08-05	1.55	123.58	-0.29%	-0.65%

z in [1.0, 2.0) · n=100 episodes · 21-session hit rate 45% · median +5d -0.01, +21d -0.11 · grade C

Episodes when this series last sat in the same trailing-365-observation z bucket, with the forward 5- and 21-session move of Broad dollar. 100 episodes at least 21 days apart; the table shows the most recent 8. A conditional distribution with its count, not a signal.

VIX · last 16.46 as of 2026-10-15 · 9,296 observations from 1990-01-02

Five years of VIX with the high-volatility weeks shaded, from the same two-state filter the Regime tab draws. One state, one shading, every tab.

The last 8 times this looked like today — and what S&P 500 did next

episode	z	level	S&P 500 +5d	S&P 500 +21d
2026-09-10	-0.13	17.84	+0.61%	+2.90%
2026-07-29	0.30	20.66	+5.57%	+5.67%
2026-07-08	-0.47	16.90	+1.20%	+3.04%
2026-06-11	0.06	19.44	+1.44%	+2.45%
2026-05-20	-0.35	17.44	+1.18%	+0.91%
2026-04-29	-0.04	18.81	+3.21%	+5.99%
2026-04-08	0.42	21.04	+3.54%	+8.17%
2026-02-25	-0.12	17.93	-1.10%	-6.75%

z in [-0.5, 0.5) · n=342 episodes · 21-session hit rate 21% · median +5d 0.00, +21d 0.00 · grade C

Episodes when this series last sat in the same trailing-365-observation z bucket, with the forward 5- and 21-session move of S&P 500. 342 episodes at least 21 days apart; the table shows the most recent 8. A conditional distribution with its count, not a signal.

What changed since last week

leaf	then	now	change	compared with
horizon days	63.00	63.00	unchanged	2026-10-11
n series	36.00	36.00	unchanged	2026-10-11
n panels	18.00	18.00	unchanged	2026-10-11

0 of 3 headline leaves moved. A repeated value can mean the upstream had not published, not that nothing happened.

ACF of the remainder · 4 rows

series	n	lags outside 95 band	first lag outside	conf95	reading	STL remainder p: Ljung-Box(24) · Jarque-Bera · ARCH-LM(4)	AutoReg (BIC lags, persistence) and its residual p, same three tests	fit seconds
10y Treasury	260	21	1	0.12	Autocorrelation of the MSTL remainder over 24 lags, with the 95% band at ±0.122,	0.000 · 0.000 · 0.000	lags none, persistence +0.00; no AutoReg residual: BIC selected no lag	0.5
WTI crude	260	18	1	0.12	Autocorrelation of the MSTL remainder over 24 lags, with the 95% band at ±0.122,	0.000 · 0.000 · 0.000	lags 1, 2, persistence -0.07; 0.057 · 0.000 · 0.000	1.4
Broad dollar	260	21	1	0.12	Autocorrelation of the MSTL remainder over 24 lags, with the 95% band at ±0.122,	0.000 · 0.000 · 0.000	lags none, persistence +0.00; no AutoReg residual: BIC selected no lag	0.4
VIX	260	4	1	0.12	Autocorrelation of the MSTL remainder over 24 lags, with the 95% band at ±0.122,	0.000 · 0.000 · 0.000	lags 1, persistence -0.11; 0.005 · 0.000 · 0.000	0.5

Harvey-Jaeger style structural check: the autocorrelation of the decomposition's remainder against a 95% band. Lags outside the band mean the trend and seasonal have not taken everything systematic out of the series. This is Method content. It is not a forecast and nothing is extrapolated from it.

cumulative abnormal return vs the sample's own mean drift, aligned by DATE to the nearest session; only the first day of each episode counts as an event; 95% CIs from 2,000 bootstrap resamples. Events are rebuilt from the underlying official series, so these are historical relationships rather than a record of calls we published. Window 10 sessions before, 20 after. As of 2026-10-18 11:56 UTC. Floor for a conditional claim: 20 episodes (grade_t_thresholds.json).

2s10s crosses into inversion — S&P 500 · n=44 · earliest 1978-08-18

horizon	n	mean	median	95% CI	t	hit rate	CI
+1d	44	+0.03%	+0.03%	[-0.2, +0.3]	0.3	55%	straddles zero
+5d	44	-0.13%	-0.25%	[-0.7, +0.4]	-0.48	46%	straddles zero
+10d	44	+0.03%	+0.41%	[-0.8, +0.8]	0.07	59%	straddles zero
+20d	44	+0.37%	+1.04%	[-0.7, +1.4]	0.66	64%	straddles zero

series · 4 rows

label	unit	kind	last	as of	first obs	n obs	ma 13w	vs 13w
10y Treasury	pct	level	5.10	2026-10-15	1962-01-02	16,182	4.91	0.19
WTI crude	usd	price	101.91	2026-10-14	1986-01-02	9,523	91.74	10.17
Broad dollar	idx	price	121.60	2026-10-09	2006-01-02	5,208	119.60	2.00
VIX	pct	level	16.46	2026-10-15	1990-01-02	9,296	15.38	1.08

scalars

field	value
horizon days	63
n series	36
statsmodels version	0.14.6
statsmodels pinned	0.14.6
statsmodels error	not reported

4 rows and 1 field withheld at this tier: absent from the page, not hidden in it.

What a 31-day filing lag is worth, measured Grade D · descriptiveas of 2026-09-11 18:01 UTC

ReadHistoryData

A fundamental exists on the day a quarter ends and becomes public on the day the filing is submitted, a median of 31 days later in this universe. A backtest that stamps the number at the period end therefore trades on it for about a third of every quarter before anyone could have seen it, and is paid for the price move that happened while the market was still waiting. Four of the five quantamental repositories this work draws on do exactly that and none of them measures it. So we run the same earnings-yield quintile long-short twice over 147 names and 155 weekly rebalances, identical in every respect but one: which vintage of the fundamental each arm may see.

cross-section on unfiled data

32%

on a typical rebalance

signal displacement

0.18 sd

cross-sectional, on the names affected

Sharpe manufactured

+0.20

and +2.4 points of return a year

arm	annualised Sharpe	95% bootstrap CI	deflated Sharpe
A: filing-stamped (what we publish)	+0.18	[-0.96, +1.26]	0.57
B: period-end-stamped (the lookahead)	+0.38	[-0.75, +1.45]	0.70

The exposure is what is measured without ambiguity: a third of the cross-section, a third of the time, priced against a filing nobody had. The Sharpe gap is the more quotable number and the weaker one, and it is reported with the test that refuses to bless it: Hansen's SPA on the difference returns p 0.15, so 155 weekly rebalances cannot separate a fifth of a Sharpe point from noise. Neither arm is traded and neither is a forecast; arm B is a deliberate error kept in the code so its size can be measured instead of assumed.

147 names, 155 weekly rebalances, 2023-09-15 to 2026-08-28. 32.1% of the cross-section on unfiled data on a typical rebalance.

The percentage of the 147-name cross-section whose fundamental had not been filed yet at each of the 60 stored weekly rebalances — the lookahead this diagnostic measures, as it actually moved. It swings with the filing calendar rather than with the market, which is the point: the error is seasonal and largest right after a quarter ends. Shaded weeks are where the site's shared four-method regime vote was split rather than unanimous — 21.7% of the 60 weeks shown.

No analog table: point 2 does not apply as written and is not forced. There is no linked market price here: both arms are the SAME long-short portfolio and the only forward quantity is the difference between them, so a "5d/21d move" would have to mean 5 and 21 weekly REBALANCES of our own diagnostic rather than sessions of anything traded. With 60 stored rebalances a 21-step forward window leaves a handful of non-overlapping episodes, under the 20 this site pre-registered as the floor for a conditional rate — so the episodes would come with a withheld hit rate and a misleading column heading. The arm-level Sharpes, their bootstrap intervals and Hansen's SPA on the difference are on the Read tab, which is the honest version of this question

What changed since last week

leaf	then	now	change	compared with
filing lag p90 days	54.00	54.00	unchanged	2026-10-11
max filing lag days	407.00	407.00	unchanged	2026-10-11
mean abs signal gap in sd	0.18	0.18	unchanged	2026-10-11
median filing lag days when arms differ	35.00	35.00	unchanged	2026-10-11
median spearman a vs b	0.98	0.98	unchanged	2026-10-11
n rebalances	155.00	155.00	unchanged	2026-10-11
n skipped	2.00	2.00	unchanged	2026-10-11
n universe	147.00	147.00	unchanged	2026-10-11

0 of 8 headline leaves moved. A repeated value can mean the upstream had not published, not that nothing happened.

Price analogs are the part of this contract that is missing here, and the reason is printed above rather than left blank.

window

key	value
from	2023-09-15
to	2026-08-28

spa

key	value
p_spa	0.30
p_lower	0.30
p_upper	0.30
t_spa	0.65
best_arm	B
null	max_k E[d_k] <= 0: no strategy in the family beats the benchmark
n_strategies	2

spa on the gap

key	value
p_spa	0.15
t_spa	1.02
null	the lookahead arm does not out-earn the honest arm

skipped · 2 rows

ticker	reason
META	no quarter with four consecutive net-income periods and a share count
SHOP	no quarter with four consecutive net-income periods and a share count

per date · 60 rows

date	n names	n vintage differs	share vintage differs	spearman a b	mean abs gap sd	ret a	ret b
2025-07-11	147	120	0.82	0.92	0.19	-0.03	-0.03
2025-07-18	147	114	0.78	0.92	0.19	0.02	0.02
2025-07-25	147	86	0.58	0.93	0.22	-0.01	-0.01
2025-08-01	147	57	0.39	0.95	0.28	0.03	0.02
2025-08-08	147	19	0.13	1.00	0.07	0.00	0.00
2025-08-15	147	16	0.11	1.00	0.06	0.02	0.02
2025-08-22	147	13	0.09	1.00	0.05	0.00	0.00
2025-08-29	147	6	0.04	1.00	0.12	-0.00	-0.000846
2025-09-05	147	6	0.04	1.00	0.09	0.00	-0.00
2025-09-12	147	5	0.03	1.00	0.08	-0.02	-0.02
2025-09-19	147	4	0.03	1.00	0.10	-0.00	-0.00
2025-09-26	147	5	0.03	1.00	0.10	-0.00042	-0.00042
2025-10-03	147	120	0.82	0.92	0.23	-0.03	-0.03
2025-10-10	147	118	0.80	0.93	0.24	-0.01	-0.01
2025-10-17	147	116	0.79	0.93	0.22	-0.02	-0.02
2025-10-24	147	89	0.61	0.93	0.24	0.00	0.00
2025-10-31	147	53	0.36	0.98	0.16	0.03	0.03
2025-11-07	147	19	0.13	1.00	0.09	0.02	0.03
2025-11-14	147	16	0.11	1.00	0.08	0.01	0.01
2025-11-21	147	11	0.07	1.00	0.03	-0.02	-0.02
2025-11-28	147	10	0.07	1.00	0.06	0.01	0.01
2025-12-05	147	8	0.05	1.00	0.08	0.03	0.03
2025-12-12	147	6	0.04	1.00	0.08	-0.01	-0.01
2025-12-19	147	1	0.01	1.00	0.04	0.01	0.01
2025-12-26	147	1	0.01	1.00	0.05	0.01	0.01
2026-01-02	147	120	0.82	0.85	0.26	0.00	0.00
2026-01-09	147	120	0.82	0.85	0.27	-0.02	0.00
2026-01-16	147	119	0.81	0.85	0.27	-0.000119	0.00
2026-01-23	147	116	0.79	0.87	0.25	0.02	0.03
2026-01-30	147	103	0.70	0.90	0.24	0.01	0.03
2026-02-06	147	94	0.64	0.90	0.25	-0.01	-0.01
2026-02-13	147	67	0.46	0.92	0.30	0.02	0.02
2026-02-20	147	44	0.30	0.97	0.22	-0.01	-0.01
2026-02-27	147	11	0.07	0.99	0.22	-0.02	-0.02
2026-03-06	147	10	0.07	0.99	0.15	-0.01	-0.01
2026-03-13	147	6	0.04	0.99	0.21	0.02	0.02
2026-03-20	147	3	0.02	1.00	0.07	0.01	0.01
2026-03-27	147	3	0.02	1.00	0.06	-0.01	-0.01
2026-04-03	147	117	0.80	0.92	0.22	0.00	0.0008
2026-04-10	147	118	0.80	0.93	0.22	-0.02	-0.03
2026-04-17	147	115	0.78	0.92	0.21	-0.04	-0.03
2026-04-24	147	91	0.62	0.93	0.23	-0.01	-0.01
2026-05-01	147	50	0.34	0.96	0.20	-0.04	-0.03
2026-05-08	147	17	0.12	0.99	0.14	-0.02	-0.02
2026-05-15	147	16	0.11	1.00	0.11	0.01	0.01
2026-05-22	147	11	0.07	1.00	0.10	-0.02	-0.02
2026-05-29	147	7	0.05	0.99	0.23	0.05	0.04
2026-06-05	147	6	0.04	0.99	0.24	-0.01	-0.01
2026-06-12	147	5	0.03	0.99	0.26	-0.01	-0.01
2026-06-19	147	4	0.03	0.99	0.34	0.02	0.01
2026-06-26	147	3	0.02	1.00	0.17	0.00	0.00
2026-07-03	147	103	0.70	0.87	0.36	0.01	0.01
2026-07-10	147	102	0.69	0.87	0.36	0.04	0.04
2026-07-17	147	99	0.67	0.87	0.36	0.02	0.03
2026-07-24	147	76	0.52	0.89	0.37	0.000637	0.02
2026-07-31	147	56	0.38	0.91	0.42	-0.03	-0.04
2026-08-07	147	15	0.10	1.00	0.18	0.00	0.01
2026-08-14	147	13	0.09	1.00	0.14	0.02	0.03
2026-08-21	147	10	0.07	1.00	0.12	-0.00	0.00
2026-08-28	147	2	0.01	1.00	0.15	-0.00	-0.00

scalars

field	value
signal	trailing twelve month earnings yield (TTM net income per share / price)
n universe	147
n rebalances	155
sharpe gap annualised	0.20
return gap pct per year	2.43
share of cross section using unfiled data	0.32
median spearman a vs b	0.98
mean abs signal gap in sd	0.18
median filing lag days when arms differ	35
filing lag p90 days	54
max filing lag days	407
pbo	not reported
price source	cache
n skipped	2

The null-result registry — everything we tested, including what failed Grade D · descriptiveas of 2026-10-15 16:11 UTC

ReadData

N 13F ownership feature 'd_managers' predicts the quarterly return AFTER the filing becomes public