Track or Follow? A Jev-Compatible Multi-Agent Trading System When Web News Lags the Market
Authors: Yebo Feng, Guangyan Gan, Liangrui Pan, Ruichao Liang, Cong Wu, Jiahua Xu, Yang Liu
Organizations: Nanyang Technological University Singapore · School of Cyber Science and Engineering Wuhan University Wuhan, China · UCL Centre for Blockchain Technologies and DLT Science Foundation London, United Kingdom
LLM agents increasingly trade on the Web news they read, yet much Web news about listed firms is published after, and about, the price move it reports, so a reader can agree with the market by following it instead of reading the event. We ask whether LLM judgments of Web news track the market's reaction to the reported event or follow it. To answer, we built a multi-agent trading system whose agents share only point-in-time stores and answer typed questions through the interface of Jev, a hosted typed decision model; code sets every position under a proved bound on how far text can move it. In a hash-committed test of a Jev-compatible reader built from a pinned LLM's option probabilities, item judgments track announcement-window abnormal returns (weekly Spearman 0.119), also after the LLM's release, and event types agree with the issuers' 8-K filings (AUC 0.833); Jev's judgments of news published after its own release track the reaction as well (descriptive). Exploratory analyses show where the association comes from: 72% of earnings news items appear after the issuer's filing, contributor-written items often report the price move, and from the first trade after publication the association falls to 0.023. At a weekly horizon, the confirmatory test detects no rank information over code-only features. Under attack, decision cards traced 88% of successful attacks to the planted item. We propose that market-graded tests of Web-text readers report the criterion from publication, not only from the event.
Figures & tables
Typed judgment
one probability per declared option, returned without generated text
emu; Jev; NG; VERB
confirmatory reader from a pinned LLM’s option probabilities; hosted typed model; open typed model; the same LLM’s verbalized probabilities
Frame
kept items whose top event type is not stock activity or opinion; H2’s units
Band
names of trusted rank at most K+J , the only names text can admit
Signal; system book
selected names at equal weight; names selected with exclusions and an incumbent buffer, sized under caps and risk rules
Figure 1. The multi-agent trading system. Agents share only point-in-time stores and answer through one interface that follows Jev’s schema; code decides under a bound that holds for any reader, and a validity harness compares the judgments with records the reader does not write. Diagram with three vertical trust lanes: untrusted Web news, semi-trusted annual-report text and trusted prices, filings and macroeconomic data. Each lane feeds a shared blackboard of point-in-time stores with a blinder. Below it sit the agents: news agents and a policy agent in the untrusted lane, a profile agent in the semi-trusted lane, and technical, fundamentals and macro agents in the trusted lane; each reads from and writes to the blackboard and answers through one typed interface that follows the Jev schema (readers Jev, emu, VERB, NG). Below, a decision layer in code (trust split, capped text logit, band, sizing, risk rules, decision card) reads the judgments and, through a separate double line that crosses the interface bar, trusted data; a tag says the bound holds for any reader. A validity harness, marked evaluation only, compares judgments with abnormal returns, 8-K items and re-identification and date probes, with nothing flowing back.
A. Confirmatory reader (emu), all 27,556 frame items
Measure
Est.
95% CI
C
AR [0,+1] , weekly mean (115 weeks)
0.119
[0.098, 0.140]
R
after the model’s release (73 weeks)
0.115
[0.090, 0.140]
R
published before 09:30 ET
0.101
[0.079, 0.124]
R
declared prior looks excluded
0.119
[0.098, 0.140]
R
first 13 weeks dropped
0.115
[0.094, 0.137]
Table 2. Validity of item judgments against the market’s reaction and the issuer’s 8-K filings: (A) the confirmatory reader, (B) the verbalized contrast. Typed item judgments track announcement-window abnormal returns (weekly Spearman 0.119, Holm-adjusted p<10−6 ). Newey–West t= 11.3.
Figure 2. Judgments track the reaction to the reported event, but mostly a reaction that precedes the item. (a) Registered estimates, (b) exploratory construct checks, (c) registered 8-K agreement. Open markers: placebo windows; light indigo: emu’s sensitivities; grey: reference readers. 95% intervals; intervals narrower than the marker are hidden. Three panels of horizontal interval plots. Panel a lists the confirmatory estimate with its registered sensitivities, all of similar size, the smaller placebo, and pooled values for the confirmatory reader and the verbalized contrast with its placebo. Panel b, shaded and labelled exploratory, shows the statistic by publication time (largest for items published during the session); from the first trade after publication it is near zero for emu and zero for two reference readers, with a small positive paired difference between emu and FinBERT; it is lower without price-report items and far higher on items that report the move; the reference readers lie below the registered estimate on the announcement window. Panel c shows AUCs against 8-K items: high for earnings results versus Item 2.02, lower for deals and management changes.
Share
Reports
ρ with
ρ from
Items
(%)
move (%)
AR [0,+1]
publication
Desk bylines
64
0.6
0.088
0.021
Named contributors
36
19.0
0.173
0.032
Before the open, event day
42
—
0.105
0.035
During the session
40
—
0.155
0.020
After the close or off-session
17
—
0.060
−0.018
Table 3. Who writes Web financial news and when it appears (exploratory, except the registered all-items AR [0,+1] value): share of frame items, share that reports the price move, and the weekly Spearman correlation ρ of E[effect] with AR [0,+1] and with the return from the first trade after publication.
Figure 3. When the criterion is realized. (a) One item’s timeline (schematic). (b) Exploratory: Spearman correlation of E[effect] with the abnormal return on each trading day around the event session, with 95% Newey–West bands; it peaks on the event session, and the day before is smaller but above zero. Panel a is a schematic time line: prior close, reported event, market reaction, item published, close of the session after the event session, then after a break the decision cutoff and the fill; a bracket marks the registered window from the prior close to that close and a dashed bracket the exploratory window from publication to the same close. Panel b is a line plot over trading days from five before to ten after the event session: the correlation is near zero before the event, rises on the day before, peaks on day zero and returns to near zero from the next day on, with shaded intervals.
Rank information at the decision horizon, 113 learned weeks
Measure
Est.
95% CI
C
H3: Δ IC over the code-only twin
− 0.016
[ − 0.029, − 0.002]
R
text part
− 0.006
[ − 0.017, 0.004]
R
trusted typed part
− 0.009
[ − 0.021, 0.002]
R
after release, 73 weeks
− 0.009
[ − 0.028, 0.011]
R
IC of the full system
− 0.024
[ − 0.061, 0.014]
Table 4. Rank information the judgments added at the decision horizon: H3 and its registered split.
A. Per attack family, per-item reader
Attack family
ASR (%)
95% CI
No def.
n
R
A1 instruction to AI readers
0.0
[0.0, 2.1] ∗
5.6
173
R
A2 fabricated fact
22.5
[16.0, 29.3]
22.5
173
R
A3 self-labelled good news
0.0
[0.0, 2.1] ∗
8.8
173
R
A4 hidden instruction
0.0
[0.0, 2.1] ∗
4.4
173
R
A5 stock promotion
4.0
[1.5, 6.8]
4.8
200
Table 5. Attack success (ASR) at the tested attacker and budget: planted facts and press releases moved names into the signal book; planted instructions did not.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Reader
Model pin
Cutoff
Release
Weeks after
emu, VERB, holistic reader and generative-analyst baseline
gpt-4.1-mini-2025-04-14
2024-06-01
2025-04-14
73
NG, also the full-frame price-report coder
flock-io/this-that-model-1.0@3d927195
undocumented
2026-09-20
none
Jev (descriptive panel)
jev-1.13.0
undocumented
2026-09-15
none
Appendix
Table A1. Readers, model pins and knowledge cutoffs. Only emu’s LLM documents a cutoff, and only emu has weeks of the evaluation window after its release.
Measure (pooled over items)
NG
95% CI
AR [0,+1]
0.108
[0.016, 0.203]
emu on the same items
0.074
[ − 0.014, 0.160]
Placebo AR [−5,−1]
0.109
[0.022, 0.193]
Agreement with emu, ρ of E[effect]
0.76
—
Agreement with emu, κ of top type
0.65
—
Earnings vs Item 2.02 (130 pos.)
0.830
[0.784, 0.872]
Appendix
Table A2. NG’s registered estimates on its 500 sampled items (pooled over items, week-block 95% intervals).
Figure A1. Look-ahead checks. (a) Registered D1: probe AUC by three-month window, filled where the registered name-permutation test gives p below 0.05; ringed: the start-rule window; grey band: the buffer after the documented cutoff; dotted line: start of the post-release weeks. (b) Exploratory: the same probe on the development slice, with the permutation null band. (c) Registered D2: share of probes whose top answer is the true firm or year, with Wilson 95% intervals, and the share of “cannot tell” answers. Three panels. Panel a plots probe AUCs by window centre from early 2024 to mid-2025, with a shaded buffer, the ringed start-rule window near one half, a run of filled windows above one half in late 2024 and early 2025, and a dotted line where the post-release window begins. Panel b, shaded and labelled exploratory, shows development-slice windows inside a grey permutation band. Panel c shows re-identification shares near zero under blinding and near one without it, for the evaluation window (circles) and the development slice (grey squares), with high shares of cannot-tell answers under blinding.
Ex. ret.
Vol.
Sharpe
MDD
TO
BE
α
System book
7.3
19.5
0.38
− 18.0
28.0
− 23
− 15.7 [ − 25.8, − 5.6]
Signal book (equal weight)
18.2
27.9
0.65
− 23.0
36.8
17
− 14.8 [ − 30.2, 0.5]
Equal-weight universe
17.4
18.0
0.97
− 18.7
3.1
—
− 0.5 [ − 0.6, − 0.3]
S&P 500 ETF
12.7
16.5
0.77
− 18.8
0.4
−∞
− 3.6 [ − 8.8, 1.6]
Vol-managed equal weight
10.5
15.8
0.66
− 19.4
3.7
− 1,063
− 3.2 [ − 9.1, 2.8]
Momentum 12-1
37.8
38.9
0.97
− 33.3
13.6
202
1.0 [ − 7.3, 9.3]
Appendix
Table A3. Registered portfolio estimates, all books. Estimates, not tests.
Exploratory (X): weekly ρ
(R)
Reader’s top
Share
AR [0,+1]
From
8-K
event type
(%)
all
no price
publ.
AUC
Product or contract
29.3
0.049 †
0.037 †
0.019
n/a
Analyst action
11.2
0.173 †
0.167 †
0.043
n/a
Earnings results
10.8
0.226 †
0.143 †
− 0.012
0.833
Other development
10.7
0.056 †
0.037
0.024
n/a
Appendix
Table A4. Where the reader tracks the reaction, by its top event type (exploratory, except the registered 8-K column). On the publication-aligned criterion, the correlation is small for every type.