Regime Watch · weekly · issue 2026-09-02 · data as of 2026-08-31 Whose web attention just changed state, and how persistent is the new state?
As of 2026-08-31, four watched names switched into state 3. AMZN and GOOG transitioned to state 3, recording staying probabilities of 0.34 and 0.36 on 1 day in state each. ORA recorded a staying probability of 0.13 on 1 day in state.
Two names moved into state 1. BA shifted to state 1 with a staying probability of 0.19 on 3 days in state, while NWBO recorded 0.09 on 3 days in state.
On the long-dwell board, names sitting past their expected dwell carry model fit flags rather than persistence. WMT posts a fit flag ratio of 33.6, BH registers a fit flag ratio of 21.6, and TAL carries a fit flag ratio of 7.7.
TAL represents 1 new entrant on the long-dwelling board.
Regime tracking utilizes forward-filtered state probabilities rather than smoothed decodes.
Model regimes measure attention level rather than price direction.
Persistence is memoryless chain math. · A regime is not a forecast of price.
Attention regime board (anchor names + this week's trend crossers)
Name
State (1 = lowest attention)
Days in state
Expected dwell (days)
P(stay 5d)
P(stay 21d)
AMZN
3
1
20.2
0.78
0.34
GOOG
3
1
20.8
0.78
0.36
ORA
3
1
10.8
0.62
0.13
WMG
2
1
9.1
0.56
0.09
MKTX
2
1
10.0
0.59
0.11
BA
1
3
13.0
0.67
0.19
NWBO
1
3
9.1
0.56
0.09
PBT
2
4
10.0
0.59
0.11
DIS
2
5
10.0
0.59
0.11
PNBK
3
6
10.0
0.59
0.11
UNH
3
10
15.9
0.72
0.26
WFC
2
13
26.4
0.82
0.44
MSFT
3
15
9.3
0.57
0.09
PG
3
18
10.0
0.59
0.11
GFS
3
22
15.3
0.71
0.24
ULBI
3
25
13.1
0.67
0.19
BDN
3
26
9.1
0.56
0.09
AVGO
2
31
78.6
0.94
0.76
OLED
3
31
9.1
0.56
0.09
TATT
3
31
13.4
0.68
0.20
TSLA
2
34
9.1
0.56
0.09
CXM
3
34
33.4
0.86
0.53
CROX
3
36
9.5
0.57
0.10
NVDA
1
41
101.8
0.95
0.81
COST
2
41
9.1
0.56
0.09
INTC
1
41
58.4
0.92
0.70
AMD
2
42
10.4
0.60
0.12
PLTR
2
53
121.2
0.96
0.84
RJET
2
55
90.4
0.95
0.79
HTZ
2
62
92.4
0.95
0.80
V
3
66
13.2
0.68
0.19
CRS
3
66
10.0
0.59
0.11
KEX
3
68
7.6
0.49
0.05
PYPL
3
75
10.0
0.59
0.11
AMBA
3
76
10.0
0.59
0.11
AAPL
1
80
42.2
0.89
0.60
JPM
3
82
8.1
0.52
0.06
ADBE
1
82
10.0
0.59
0.11
HD
3
97
13.6
0.68
0.20
TAL
3
100
13.0
0.67
0.19
A hidden-Markov regime model on each name's own attention series (states ranked 1 = lowest). Persistence is memoryless chain math — sitting modestly past the expected dwell is normal; a Days ÷ expected dwell above 5 is a fit flag, not persistence. Descriptive regime measurement, not a forecast of attention or price. States (K): 3 (all rows). As of: 2026-08-31 (all rows).
No change in the set since 2026-09-01 (all 10 carried over).
Switched state in the last 7 days
Name
New state
Days since switch
Expected dwell (days)
P(stay 21d)
AMZN
3
1
20.2
0.34
GOOG
3
1
20.8
0.36
ORA
3
1
10.8
0.13
WMG
2
1
9.1
0.09
MKTX
2
1
10.0
0.11
BA
1
3
13.0
0.19
NWBO
1
3
9.1
0.09
PBT
2
4
10.0
0.11
DIS
2
5
10.0
0.11
PNBK
3
6
10.0
0.11
Empty when no watched name changed state this week — that is a finding, not a gap. States (K): 3 (all rows).
Long-dwelling states (40 or more sessions in one state)
Name
State
Days in state
Expected dwell (days)
Days ÷ expected dwell
WMT
1
567
16.9
33.6
BH
3
294
13.6
21.6
XOM
3
152
9.3
16.3
NFLX
3
184
11.3
16.3
VRT
3
154
10.0
15.4
GEHC
3
135
10.4
13.0
COIN
2
126
10.0
12.6
AGCO
3
105
8.6
12.2
JPM
3
82
8.1
10.1
KEX
3
68
7.6
8.9
ADBE
1
82
10.0
8.2
TAL
3
100
13.0
7.7
A ratio above 5 means the fitted chain does not describe that name's current state — a fit flag, not persistence. Flagged, not explained.
AMZN · web attention as a fraction of market value, 1 year; dated marks = regime switches and front-page stories
GOOG · web attention as a fraction of market value, 1 year; dated marks = regime switches and front-page stories
ORA · web attention as a fraction of market value, 1 year; dated marks = regime switches and front-page stories
Split reading: 1 of 3 mixed, 1 of 3 continuation, 1 of 3 rotation.
House view · panel order (no scored history yet)
Interpretation · anthropic/claude-opus-5 · data as of 2026-08-31 · published 2026-09-02 · a view, not a cited fact; recorded so it can be scored
This week's switch board is two events wearing one label: mega-cap attention in AMZN and GOOG has genuinely re-levelled into state 3 on slow chains, while the short-dwell crossers around them are day-to-day churn that will not carry. AMZN (0.34) and GOOG (0.36) sit on roughly 20-day expected dwell; ORA's 0.13 and the 9-10 day cohort describe attention that reverts before a reader can use it. WMT's 33.6 ratio is a fit flag, not staying power. We expect both AMZN and GOOG to still carry state 3 a week from now.
What would show this wrong: Either AMZN or GOOG shows a state other than 3 on the regime board in the issue one week after 2026-08-31.
Population: runs of consecutive sessions in one attention state on each watch-list name's state path (the forward-filtered decode, which labels each day with data to that day only, on every vintage fit from 2026-09-02 -- 51 of 51 names here; earlier vintages hold only the smoothed full-series decode, whose labels use later data and carry fewer short runs, so a rate mixing the two leans toward persistence; the model is refit daily on the full history, the path keeps the last 500 sessions; states ranked 1 = lowest attention). The final run of every path is right-censored (still running at the vintage date) and is excluded from the rate; the first run began before the path starts. AMZN: of 249 completed runs of state 3 (of 3) that reached day 1, 73% reached day 6. GOOG: of 249 completed runs of state 3 (of 3) that reached day 1, 73% reached day 6. Joint rate 53%: per-name rates multiplied as if independent (they are not). Window 2021-08-11 to 2026-08-31. Base rate 53% (conditional, n=249); this view is scored as lift over that rate. Resolves by 2026-09-09. Claim a05bf50b.
Other readings (2)
Interpretation · google/gemini-3.8-flash · data as of 2026-08-31 · published 2026-09-02 · a view, not a cited fact; recorded so it can be scored
Elevated web attention around GOOG reflects a durable shift into top-tier interest rather than a brief spike. With an expected dwell of 20.8 days in state 3, crowd engagement is settling into an extended period of scrutiny. GOOG will remain in attention state 3 through the next two weeks.
What would show this wrong: GOOG switches out of state 3 within two weeks.
Population: runs of consecutive sessions in one attention state on each watch-list name's state path (the forward-filtered decode, which labels each day with data to that day only, on every vintage fit from 2026-09-02 -- 51 of 51 names here; earlier vintages hold only the smoothed full-series decode, whose labels use later data and carry fewer short runs, so a rate mixing the two leans toward persistence; the model is refit daily on the full history, the path keeps the last 500 sessions; states ranked 1 = lowest attention). The final run of every path is right-censored (still running at the vintage date) and is excluded from the rate; the first run began before the path starts. GOOG: of 249 completed runs of state 3 (of 3) that reached day 1, 57% reached day 11. Window 2021-08-11 to 2026-08-31. Base rate 57% (conditional, n=249); this view is scored as lift over that rate. Resolves by 2026-09-16. Claim d617fe63.
Checked against the tables (1)
'Elevated web attention around GOOG reflects a durable shift into top-tier interest rather than a brief spike.' — compares rows the tables do not rank that way (top).
Interpretation · anthropic/claude-fable-5.1 · data as of 2026-08-31 · published 2026-09-02 · a view, not a cited fact; recorded so it can be scored
AMZN and GOOG entered the top attention state on the same day (1 day in state each), with staying probabilities of 0.34 and 0.36 and an expected dwell of 20.2 days for AMZN; ORA's parallel move carries only 0.13. We read the paired mega-cap switch as a broad tech attention wave rather than one-off headlines, and such waves outlast single-name spikes. BA and NWBO sliding to state 1 says the crowd is concentrating on mega-cap tech, not broadening. AMZN remains in state 3 through the next two weeks. The long-dwell flags on WMT and BH are fit problems, not regimes.
What would show this wrong: AMZN's forward-filtered state is reported as 1 or 2 in any issue dated on or before 2026-09-14.
Population: runs of consecutive sessions in one attention state on each watch-list name's state path (the forward-filtered decode, which labels each day with data to that day only, on every vintage fit from 2026-09-02 -- 51 of 51 names here; earlier vintages hold only the smoothed full-series decode, whose labels use later data and carry fewer short runs, so a rate mixing the two leans toward persistence; the model is refit daily on the full history, the path keeps the last 500 sessions; states ranked 1 = lowest attention). The final run of every path is right-censored (still running at the vintage date) and is excluded from the rate; the first run began before the path starts. AMZN: of 249 completed runs of state 3 (of 3) that reached day 1, 57% reached day 11. Window 2021-08-11 to 2026-08-31. Base rate 57% (conditional, n=249); this view is scored as lift over that rate. Resolves by 2026-09-16. Claim 310ffb94.
Checked against the tables (2)
'AMZN and GOOG entered the top attention state on the same day (1 day in state each), with staying probabilitie' — compares rows the tables do not rank that way (top).
'BA and NWBO sliding to state 1 says the crowd is concentrating on mega-cap tech, not broadening.' — a claim about a group (concentrating) the tables do not aggregate.
Known open data defects affecting this publication: Digital-asset market caps are present but stale: the upstream cap feed froze silently. (2026-02-17 to 2026-08-04); A name whose attention falls below the publication floor silently leaves the published universe; the exit is indistinguishable from a delist (always); The historical cross-section is today's universe backfilled: a name onboarded later carries rows for earlier dates. 115 of 2,135 names end b (2015-05 to 2026-08). Coverage & data quality
Method
Each week the same hidden-Markov regime model is fitted to the web-attention series of a fixed watch list — two dozen anchor names plus every name that crossed its 50-day attention trend in the last seven sessions. The model picks two or three attention states by information criterion, ranks them by level, and reports for each name the current state, how long it has been there, the expected dwell, and the probability of remaining in the state over 5 and 21 sessions. The switch table lists names whose current state began within the last seven days; the long-dwell table lists names that have sat in one state for well over the model's expected dwell.
Regimes describe the level and persistence of attention, not its cause and not any price outcome. Persistence probabilities are memoryless chain math, so a state outlasting its expected dwell by a factor of two is ordinary. A state that has lasted more than five times the expected dwell is not: it means the fitted chain does not describe that name's current state, and the long-dwell table flags the ratio as a fit problem rather than as persistence. No return claim attaches to any table on this page.
From 2026-09-02 the state on the board is the forward-filtered one: the probability of each state given the series up to that day, so the label a name carries is the label a reader could have known that day. The smoothed decode that labels each day with the whole series — the one a model fit prints by default — is kept as a research view; it is smoother, has fewer short runs, and cannot be the basis of a persistence claim or its base rate, because a switch it reports on a Tuesday may only be visible from the Friday. The two decodes agree on most days for most names; the agreement share is a hidden column on every row. The parameters are still refit daily on the full history.
Published by AssetFlow.ai LLC. Descriptive research on attention and public-record data; not investment advice and not an offer or solicitation. No return claim attaches to anything above. Issue date 2026-09-02.