Thinking Fast, Trading Late: System One Decision Models vs the Market

Amir Dehkordi; Alireza Kargarzadeh; Nariman Khaledian; Navid Parvini; Sid Ghatak; Arman Khaledian

cs.AINov 2, 2026

Thinking Fast, Trading Late: System One Decision Models vs the Market

Authors: Amir Dehkordi, Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini, Sid Ghatak, Arman Khaledian

Organizations: Independent Researcher · TailState Intelligence Ltd · Zanista AI Ltd · Increase Alpha, LLC

Abstract

Large language models (LLMs) read company news at scale, and backtests suggest that their readings predict stock returns. We ask how much of that predictability survives a test that rules out look-ahead, memorization and selection. We propose OLIVIA, a protocol for such tests: point-in-time inputs with company identities masked, selection on validation data only, an analysis plan frozen by cryptographic hash, and a test period unblinded once through a logged lock. We apply it to AI decision models, non-generative language models that return probabilities for typed questions, reading news about S&P 500 stocks: an open-weights model we fine-tuned on 2024-2025 data, three commercial models and standard baselines. The fine-tuned model's signal replicated on two validation quarters (daily information coefficients 0.055 and 0.065) but fell to 0.020 (t = 1.7) on the blinded January-July 2026 test. The commercial models' readings line up with the overnight move to the open (information coefficient 0.074 to 0.077) but not with the move after it, so the news is priced by the time a trader can act on it at the open. No model can tell over- from under-reaction, and no pre-registered strategy survives costs and a deflated Sharpe ratio. Commercial models identify masked companies in 62% to 74% of cases against a 28% baseline: masking hides names, not companies. The market reads the news first.

Figures & tables

Safeguard	Guards against	In this study	A replication needs
O ne-time unblinding	Repeated looks at the test period	The 2026 prices were sealed behind a lock that checks the frozen hash; it released the test outcomes only after the freeze, to the frozen analyses, and logs every read	Sealed outcomes and a log of every access
L ook-ahead-free states	Inputs that were not public at the decision	Every input carries the time it became known; the news ends at 09:15 ET, a rounded time stamp counts from the end of its hour or day, and rolling statistics end the day before	An availability time for every input
I dentity masking	A model recalling what happened to a company it recognizes	Names, tickers, people, dates and price moves are removed; an independent scanner checks the result; a matched-distractor test measures recognition (Section )	A masker, a leak check and a recognition test
V alidation-only selection	Choosing models, filters or thresholds on the test	Search and calibration on val-A; confirmation and the final selection on val-B	Two validation periods before the test
I mmutable plan	Changing hypotheses or tests after seeing results	The pre-registration and the analysis code were frozen by a SHA-256 hash and a version-control tag before any test-period call; later changes go to a dated, append-only deviations log	A hash, a tag and a deviations log
A udit trail	Unreported variants	Append-only ledgers of every evaluated variant (their count sets the bar of the deflated Sharpe ratio), of API spend and of every access to test-period data	A trial ledger kept from the first analysis

Table 1: The OLIVIA protocol: six safeguards for testing language-model signals on financial news, what each guards against, how this study implements it and what a replication needs. The initials of the safeguards give the name.

Panel of Figure 1 — Figure 1: The design: every choice was made on 2025 validation data, the 2026 test was sealed until the plan was frozen, and every model was released after the sample ended. (a) The calendar; the time axis changes scale at the break, from months to days. The test outcomes were released only after the freeze. The forward test, pre-registered on 2026-10-29 and not yet run, uses news published from 2026-10-22 on. (b) One trading session. At D0 (09:15 ET) a model reads the news published since the previous close; at D1 (10:00) and D2 (after the close) it also sees the reaction so far. The overnight move $z_{on}$ runs from the previous close to the open, so it includes after-hours and pre-market trading; the move after the open $z_{id}$ runs from the open to the close; and $y_{2}$ runs from the next open to the close of $t+5$ . The convergence ladder of Section runs to $C_{t+20}$ .

Figure 2: Every validation IC that was clearly positive fell on the blinded test: Laya-Fin’s from 0.065 to 0.020, d1-ultra’s from 0.097 to 0.039. Daily IC of the pre-open reading

\hat{N}

with the close-to-close move

z_{cc}

on (a) val-B (October–December 2025, the final-selection quarter) and (b) the test period (January–July 2026), with 95% intervals (

\pm 1.96

Newey–West standard errors); the panels share one scale. Colors mark the series that recur across the figures: Laya-Fin (blue), d1-ultra (orange) and the price-only benchmark (green). The LLM baseline was run on a 2,000-event subsample. † Vendor-label baselines: the labels were produced after the fact, so their ICs are a possibly look-ahead-contaminated upper bound. ‡ See Appendix .

Hypothesis	Estimate	95% CI	$p$	Holm $p$	Estimate	$p$	Verdict
	Laya-Fin (primary)				Anchor
H1	$0.005$	[ $-0.035$ , $0.041$ ]	$0.409$	$0.853$	$0.004$	$0.406$	not supported
H1b	$0.010$	[ $-0.014$ , $0.037$ ]	$0.213$	$0.853$	$0.002$	$0.427$	not supported
H2	$-0.044$	—	$0.616$	$0.853$	$0.015$	$0.294$	not supported
H3	$0.0038$	[ $-0.0008$ , $0.0085$ ]	$0.054$	$0.269$	$0.0013$	$0.017$	not supported
H4	$0.0012$	[ $-0.0017$ , $0.0042$ ]	$0.216$	$0.853$	$-0.0004$	$0.773$	not supported

Table 2: Primary hypotheses on the test period: none of the five is supported for the primary instrument, Laya-Fin (Holm

p

from 0.269 to 0.853).

Figure 3: The commercial models’ pre-open readings line up with the overnight move (daily IC 0.074 to 0.077), not with the move after the open (about zero). Daily IC of the pre-open reading

\hat{N}

on the test period with (a) the overnight move

z_{on}

(previous close to open) and (b) the move after the open

z_{id}

(open to close), with 95% bootstrap intervals; (c) H1, the difference of the two, with its bootstrap interval. The panels share one scale. Laya-Fin’s ICs are small on both sides (0.015 and 0.010), so H1 and H1b are not supported for it (Holm

p=0.853

). ‡ See Appendix .

	Val-B			Test
Model	$z_{cc}$	$z_{on}$	$z_{id}$	$z_{cc}$	$z_{on}$	$z_{id}$
Trained and zero-shot
Laya-Fin	$0.065$	$0.045$	$0.054$	$0.020$	$0.015$	$0.010$
	( $3.64$ )	( $2.93$ )	( $2.89$ )	( $1.67$ )	( $1.08$ )	( $0.82$ )
anchor	$-0.020$	$-0.002$	$-0.011$	$0.008$	$0.006$	$0.002$
	( $-1.52$ )	( $-0.16$ )	( $-0.75$ )	( $0.83$ )	( $0.50$ )	( $0.18$ )

Table 3: Daily IC by target: the API models’ IC is 0.074 to 0.077 with the overnight move and about zero with the move after the open; no model predicts the move after the open.

Decision point	$b$	SE	$p_{b}$	$c$	SE	$p_{c}$	Events
Laya-Fin (primary)
D2, close (H2)	$-0.044$	( $0.149$ )	$0.616$	$-0.015$	( $0.012$ )	$0.110$	$17,188$
D1, 10:00	$0.229$	( $0.148$ )	$0.062$	$0.013$	( $0.012$ )	$0.861$	$13,203$
D0, open	$0.278$	( $0.212$ )	$0.096$	—	—	—	$13,439$
Anchor (uninformative)
D2, close (H2)	$0.748$	( $1.377$ )	$0.294$	$-0.015$	( $0.012$ )	$0.111$	$17,188$

Table 4: The over/under regression (H2): for Laya-Fin the news coefficient

b

-

0.044 and the reversal coefficient

c

-

0.015 (

p

= 0.110):

c

has the predicted sign but is not significant, and

b

is not positive at the close.

Figure 4: Prices move part of the way to the model’s news-implied value by the close, then partly fade: the convergence slope is 0.36 at the close for Laya-Fin and 0.59 for d1-ultra, and the gap’s half-life is not reached within 20 sessions. Test period. Event time is not to scale:

O_{t}

is the open, 10:00 the D1 decision point,

C_{t}

the close of day

t

and

+k

the close of session

t+k

. (a) d1-ultra’s top and bottom deciles of the news-implied value

V

: the level move

\mathrm{CAR}^{\mathrm{lvl}}

from the previous close, in typical daily moves, with

\pm 1.96

event-level standard-error bands (descriptive), against each decile’s mean

V

(dashed). (b) The convergence ladder

\kappa_{h}

of Equation ( ), with 95% intervals from standard errors clustered by date and firm, against full convergence at

\kappa_{h}=1

Figure 5: The models’ “underreaction” calls never catch up, and moves of three or more typical days reverse slightly in every tercile of news strength (

-

0.09 to

-

0.18 typical moves over the next week). Test period, d1-ultra. (a) D2 events by the post-reaction class, pooled over the size tiers and signed in the news’ direction (

\mathrm{sign}(\hat{N})\cdot\mathrm{CAR}^{\mathrm{lvl}}

), with

\pm 1.96

event-level standard-error bands, against each class’s mean signed value

\bar{V}

(dashed: dotted for Under, short dashes for Fair, long dashes for Over). (b) Drift after the close,

\mathrm{sign}(G)\cdot y_{2}

(next open to the close of

t+5

, in the direction of the day’s move), by the size of the day’s move

|G|

, one line per tercile of news strength

|V|

, with 95% intervals; for moves of three or more, two of the three intervals include zero. The strong tercile also reverses slightly for moves of 0.5 to 2 typical days. The pre-registered regression (Table ) finds no support for H2.

Figure 6: Same-day news explains 0.25% of the stock-specific daily variance out of sample for Laya-Fin and 0.82% for d1-ultra; reading the realized move explains 15.8%. Test period; out-of-sample

R^{2}

in percent of the stock-specific variance (the market and the sector removed). (a) d1-ultra’s

R^{2}

by cut, stacked by step in the legend’s order: the pre-open news, plus the day’s news, plus the post-reaction explanation (which reads the realized move, so it is not a forecast). Each total stands above its bar with its 95% bootstrap interval; the two thin news steps are labeled beside each bar. For comparison, the market and the sector explain 25.3% of the total daily variance, a different base. (b) The news steps on their own scale, with 95% bootstrap intervals, for Laya-Fin and d1-ultra. Only d1-ultra answers the explanation question.

Step	$R^{2}$	$\Delta R^{2}$	$R^{2}$	$\Delta R^{2}$	$R^{2}$	$\Delta R^{2}$
	Laya-Fin		d1-ultra		Anchor
Panel A: the explanation ladder, out-of-sample $R^{2}$ of $z_{cc}$ (%)
+ ex-ante news (D0)	$0.13$	—	$0.28$	—	$0.01$	—
+ same-day news	$0.25$	$0.12$	$0.82$	$0.53$	$0.03$	$0.02$
+ explanation (reads the move)	—	—	$15.76$	$14.94$	—	—
share knowable before the open (%)	$53.17$	—	$34.78$	—	$35.63$	—

Table 5: Explaining realized moves: same-day news explains 0.25% of the stock-specific variance for Laya-Fin (H5, supported) and 0.82% for d1-ultra; reading the move explains 15.8%.

Figure 7: Commercial decision models identify the company behind masked news in 62% to 74% of states, against a matched baseline of 28%; the open-weights Laya model does so in 32%. Company identification on masked val-A states with activity-matched distractors: the share of states in which each model names the company among four candidates, sorted, with its Wilson 95% interval, against the matched baseline (a guess that uses no reading; solid line) and the pre-registered screening threshold (40%; dashed line). The three commercial models carried to the test are in the comparator color, the other commercial models and Laya zero-shot in gray. Every model was released after the sample ended. Identifying a company is not the same as remembering its returns. ‡ See Appendix .

Laya-Fin (primary)
		Sharpe ratio
Book	Holding	Gross	Net	Turnover	Break-even (bp)	Placebo pct.	DSR
S-CONV	D2 to $t{+}5$	$0.61$	$-1.98$	$0.90$	$2.4$	$68$	$<$ $0.001$
S-D0	open to close	$0.35$	$-6.03$	$4.36$	$0.5$	$64$	$<$ $0.001$
S-D0, gated	open to $t{+}1$	$1.88$	$-0.38$	$2.31$	$8.3$	$93$	$<$ $0.001$
S-D1 $J$	10:00 to close	$0.76$	$-9.70$	$4.36$	$0.7$	$71$	$<$ $0.001$

Table 6: Strategies on the test period: no pre-registered book survives costs and the deflated Sharpe ratio (the largest deflated-Sharpe probability is 0.024).

Figure 8: No pre-registered strategy survives costs: net of 10 basis points every main book loses, and the convergence strategy is effectively a price-only gap fade (92% of positions shared, daily correlation 0.96). Test period. (a, b) Annualized Sharpe ratio of each main book, gross and net of 10 basis points per side on both legs and the SPY hedge, at the exit of its hypothesis’ horizon; the gray box is the 5th to 95th percentile of 100 random-signal books on the same events; (b) has its own scale. (c) The convergence strategy’s cumulative net return (Laya-Fin and d1-ultra) against the price-only gap fade, with the placebo’s 5th to 95th percentile band.

Figure 9: Exploratory (post hoc): the overnight signal concentrates in d1-ultra’s most confident tenth of pre-open readings in all three periods (mean signed overnight move 1.52, 1.16 and 0.93 typical moves, against 0.22, 0.26 and 0.21 for all events); the move after the open shows no such concentration. Coverage curves: events ranked by conviction

|\hat{N}|

(lines) or by the stated confidence (circles), keeping the most confident share, with thresholds from val-A; 95% intervals from the stationary bootstrap over sessions. (a) Mean

\mathrm{sign}(\hat{N})\cdot z_{on}

; (b) mean

\mathrm{sign}(\hat{N})\cdot z_{id}

; (c) the hit rate against

z_{cc}

. The dashed line marks the frozen confidence gate (the val-A 70th percentile). ‡ See Appendix .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
D0, D1, D2	Decision points: 09:15 ET before the open (pre-open reading); 10:00 ET after the first 30-minute bar and after the close (post-reaction readings).
val-A, val-B, test	2025-07-01 to 2025-09-30; 2025-10-01 to 2025-12-31; 2026-01-02 to 2026-07-31.
$r^{cc},r^{on},r^{id}$	Close-to-close, overnight (close to open) and intraday (open to close) returns, with $1+r^{cc}=(1+r^{on})(1+r^{id})$ ; Equation ( ).
$\mathrm{AR}^{k}$ , $\sigma^{k}_{i,t-1}$	Abnormal return (market and leave-one-out sector removed, betas over 252 sessions to $t-1$ ); its ex-ante scale (1.4826 times the median absolute deviation over 126 sessions to $t-1$ , floored at the day’s cross-sectional 10th percentile).
$z^{k}$ ; $z_{cc},z_{on},z_{id}$	$\mathrm{AR}^{k}/\sigma^{k}_{i,t-1}$ , the move in typical daily moves (clipped at $\pm 10$ ).
$y_{0},y_{1},y_{2}$	Outcomes after D0, D1 and D2: $z_{id}$ ; the move from 10:00 to the close; the move from the next open to the close of $t+5$ , each over its ex-ante volatility.

Appendix

Table A1: Notation.

Model	IC val-B	$t$	IC test	$t$	Events	USD/1,000	p50 (ms)
Laya-Fin	$0.065$	$3.64$	$0.020$	$1.67$	$13,748$	$0.000$	—
anchor	$-0.020$	$-1.52$	$0.008$	$0.83$	$13,748$	$0.000$	—
Laya zero-shot	$0.029$	$1.94$	$0.005$	$0.37$	$13,748$	$0.000$	$536$
d1-ultra	$0.097$	$5.38$	$0.039$	$3.23$	$13,748$	$0.063$	$288$
span-02 ‡	$0.099$	$5.83$	$0.041$	$3.10$	$13,748$	$0.028$	$2,262$
drex-v2.0	$0.075$	$4.10$	$0.044$	$3.70$	$13,748$	$0.045$	$680$

Appendix

Table A2: Model comparison: every val-B IC with

z_{cc}

that is clearly positive falls on the test (Laya-Fin 0.065 to 0.020); the cost and the latency.

Figure A1: No pre-registered robustness variant changes a conclusion: Laya-Fin’s H1 and H1b stay within their intervals of zero, and d1-ultra’s H1 stays positive in every variant, from 0.022 when the latest article precedes 04:00 to 0.122 when it arrives between 08:00 and 09:15. Test period, D0 events. Each variant raw (filled) and with placebo-adjusted targets (hollow), below the main estimate (diamond, from the stationary block bootstrap), with 95% intervals (

\pm 1.96

Newey–West standard errors for the variants). (a) and (d), both H1, share a scale; (b) and (c) have their own.

Model, cut	Events	Wald $F$	$p$	BH $q$	$\Delta b$	SE	$\Delta c$	SE
Laya-Fin, Top 100	$4,516$	$1.83$	$0.164$	$0.741$	$0.490$	( $0.281$ )	$-0.025$	( $0.021$ )
Laya-Fin, 101 to 250	$5,121$	$0.10$	$0.904$	$0.904$	$-0.008$	( $0.258$ )	$-0.007$	( $0.016$ )
Laya-Fin, 251 to 500	$7,551$	$1.71$	$0.185$	$0.741$	$-0.326$	( $0.253$ )	$0.023$	( $0.016$ )
Laya-Fin, quantified	$15,030$	$0.66$	$0.521$	$0.833$	$0.470$	( $0.609$ )	$0.022$	( $0.026$ )
Laya-Fin, small move	$9,795$	$0.29$	$0.751$	$0.859$	$-0.193$	( $0.255$ )	$0.005$	( $0.035$ )
Laya-Fin, press wire	$9,655$	$0.75$	$0.477$	$0.833$	$-0.058$	( $0.330$ )	$-0.024$	( $0.020$ )

Appendix

Table A3: Heterogeneity of H2: no cut differs (Laya-Fin’s Wald

p

from 0.16 to 0.90).

Model, test	Estimate	95% CI	SE	$p$	BH $q$	Pooled $q$
Laya-Fin, H5	$0.0025$	—	—	$0.010$	$0.058$	$0.078$
Laya-Fin, H6	—	—	—	—	—	—
Laya-Fin, H7	$0.0146$	[ $-0.0068$ , $0.0363$ ]	—	$0.097$	$0.238$	$0.317$
Laya-Fin, H2-C, $\kappa$ at $h^{*}$	$0.2588$	—	( $0.2299$ )	$0.131$	$0.238$	$0.317$
Laya-Fin, H2-C, $\kappa$ increase	$0.0182$	[ $-0.3614$ , $0.4903$ ]	—	$0.467$	$0.561$	$0.748$
Laya-Fin, H2-C, $\lambda_{t+5}$	$0.0265$	—	( $0.0264$ )	$0.159$	$0.238$	$0.317$

Appendix

Table A4: Secondary hypotheses: only H5 survives Benjamini–Hochberg for Laya-Fin (

q

= 0.058).

	$O_{t}$	10:00	$C_{t}$	$t{+}1$	$t{+}2$	$t{+}5$	$t{+}10$	$t{+}20$	Half-life
Panel A: convergence ladder $\kappa_{h}$ (D0 events)
Laya-Fin	$0.24$	$0.27$	$0.36$	$0.36$	$0.38$	$0.26$	$0.21$	$0.26$	—
	( $0.07$ )	( $0.08$ )	( $0.09$ )	( $0.10$ )	( $0.11$ )	( $0.14$ )	( $0.18$ )	( $0.23$ )	—
d1-ultra	$0.55$	$0.61$	$0.59$	$0.54$	$0.47$	$0.36$	$0.25$	$0.36$	—
	( $0.08$ )	( $0.09$ )	( $0.08$ )	( $0.09$ )	( $0.11$ )	( $0.13$ )	( $0.15$ )	( $0.22$ )	—
anchor	$0.08$	$-0.02$	$-0.04$	$-0.04$	$-0.03$	$-0.06$	$-0.03$	$0.08$	—

Appendix

Table A5: Convergence to the news-implied value

V

: prices reach part of

V

on day

t

(

\kappa

at the close 0.36 for Laya-Fin and 0.59 for d1-ultra), then partly fade; the gap’s half-life is not reached within

t{+}20

Model, target	ECE	Log-loss skill	Brier skill	AUC	$T$	ECE before	ECE after
Laya-Fin, $e_{5}$	$0.057$	$0.001$	$0.000$	—	—	—	—
Laya-Fin, $z_{cc}$	$0.033$	$0.010$	$0.010$	—	—	—	—
Laya-Fin, $\|z_{cc}\|\geq 1$	—	—	—	$0.60$	—	—	—
Laya-Fin, $e_{5}$ , temperature scaling	—	—	—	—	$1.00$	$0.058$	$0.057$
anchor, $e_{5}$	$0.048$	$0.002$	$0.001$	—	—	—	—
anchor, $z_{cc}$	$0.016$	$0.015$	$0.015$	—	—	—	—

Appendix

Table A6: Calibration of the three-class forecasts: the trained models are calibrated but barely informative (Laya-Fin’s ECE 0.057), the API models overconfident (ECE 0.209 to 0.361).

Figure A2: Laya-Fin’s three-class forecasts stay close to the base rates (expected calibration error 0.057); the commercial models are overconfident (0.209 to 0.361). Reliability of the pre-open reading’s three classes (a fall, no move or a rise, split at

\pm 0.5

typical moves) against the five-day move

e_{5}

on the test period: the hit rate of the top class by bin of its stated probability (15 bins; the dot area grows with the number of events in the bin; the diagonal is perfect calibration). The anchor’s near-constant readings sit in two bins. ‡ See Appendix .

Figure A3: Exploratory (post hoc): none of the 8 conditions selected on two validation quarters held on the test; 5 changed sign. For each (model, condition) pair, the difference inside minus outside the condition of the mean

\mathrm{sign}(\hat{N})\cdot z_{id}

(typical daily moves) on val-A (circles), val-B (squares) and the test (dots), with 95% intervals; shaded rows were selected by the rule fixed before the test run (positive with

t\geq 1

in both validation quarters and a pooled

t\geq 2

). † Vendor fields are retrospective labels and only split events descriptively. “(seen)” marks conditions whose test values the authors’ assistant had seen before the study (Section ); none of them was selected.

Financial Reasoning in Language Models

See all 25 papers

Sep 20, 2026

Financial Language Models as Applied Artificial Intelligence Systems for News-Based Trading under Market Frictions

Financial language models can transform unstructured firm-specific news into structured decision signals, but financial AI research lacks an integrated deployment framework for evaluating whether those signals remain useful in financial decision systems. Computer science research has developed strong methods for time-series forecasting, text classification, multimodal stock prediction, graph-based market modeling, and machine-learning operations, yet these streams do not provide a domain-specific protocol that jointly tests financial language-model outputs under event-time observability, probability calibration, execution timing, transaction costs, liquidity constraints, capacity limits, operational diagnostics, and statistical inference. We introduce MFAST, a Market-Friction-Aware Sentiment-to-Trading framework that converts timestamped financial text into auditable, reproducible, and market-feasible trading decisions. The application is news-based trading, where firm-specific text must be linked to securities before portfolio decisions can be evaluated. The framework links Refinitiv News Analytics to Center for Research in Security Prices (CRSP) equity data, restricts the primary out-of-sample evaluation to post-release news outside disclosed foundation-model data-freshness periods, and adds a public replication arm using open financial text and public price data. Results show that decoder-only language models outperform encoder baselines and dictionary sentiment in classification, calibration, return prediction, and net portfolio performance, while operational diagnostics reveal trade-offs among accuracy, latency, memory, throughput, and inference cost. The paper shows that credible evaluation of financial language models requires an end-to-end engineering approach combining language understanding, temporal discipline, market-friction-aware deployment, and reproducible validation.

Kemal Kirtac

Department of Computer Science, University College London, 66–72 Gower Street, London WC1E 6EA, United Kingdom · Department of Computer Science, University of Warwick, 6 Lord Bhattacharyya Way, Coventry CV4 7EZ, United Kingdom

Financial Language Model Evaluation LLM Evaluation Financial Sentiment Analysis Algorithmic Trading

Ask PDF

Oct 19, 2026

Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

Large language model (LLM) agents now trade real capital using news as evidence, yet they cannot reliably tell fabricated headlines from real news. Because they read the same feeds under near-identical policies, a single fabricated post can push them into the same trade at the same moment. We present a systematic attack benchmark for information injection against LLM trading agents. Unlike prompt injection, which targets the model's instructions, information injection targets the evidence the agent trusts. The benchmark spans five injection vectors and seven event categories, evaluated over ~10,000 decisions on 15 U.S. stocks. Five findings emerge. (i) Prompt injection achieves 14% attack success rate (ASR), but fabricated news reaches 62% and two 30-word templates reach 100% on all 15 stocks, inverting the usual LLM-security intuition. (ii) For fact-poisoning templates, surface form matters more than content, since removing credibility cues such as a named source or a "BREAKING" tag drops ASR from 100% to 0% while the fake claim remains identical. (iii) Success depends on the agent's state, with a fake indicator panel scoring 91% against an agent about to hold but 12% against one about to buy. (iv) Attacks generalize across model vendors and remain effective when blended, unlabeled, into real news feeds. (v) Most attacks evade standard defenses, since model scaling stops shallow vectors but not fact poisoning, peer debate fails, and only hierarchical aggregation reduces ASR.

Mengxiao Wang, Nitesh Saxena

Texas A&M University

Adversarial Attacks on LLMs AI Agent Security LLM Security LLM Agents

Ask PDF

Jun 10, 2026

Can News Predict the Market? Limits of Zero-Shot Financial NLP and the Role of Explainable AI

Can financial news reliably predict short-term stock movements? Despite advances in large language models, this question remains unresolved. We revisit this problem using a zero-shot natural language processing framework, investigating whether models can extract actionable signals from financial news without domain-specific training. We design a structured pipeline that combines zero-shot natural language inference with temporal aggregation, explicitly modelling recency and event-dependent impact horizons when integrating information across articles. To address the need for transparency in high-stakes settings, we introduce a multi-layered explainability framework that links predictions to token-level, article-level, and aggregate evidence, and produces grounded natural language rationales. Across multiple models and prediction horizons, we find that zero-shot approaches consistently fail to outperform simple baselines, with particularly weak performance on negative movements, suggesting deeper structural limitations in mapping news sentiment to short-term price dynamics. However, explainability signals reliably distinguish between trustworthy and unreliable predictions, offering practical value even when accuracy is limited. These findings highlight the limits of zero-shot financial NLP and motivate a shift toward decision-support systems that prioritise transparency and uncertainty awareness. Code: https://github.com/alimert05/zero-shot-stock-xai

Ali M Karaoglu, Shreyank N Gowda

School of Computer Science, University of Nottingham, Nottingham, United Kingdom

Financial Forecasting Zero-Shot Text Classification Finance Explainable Artificial Intelligence Financial Sentiment Analysis Time Series Forecasting+1

Ask PDF

Abstract

Financial Reasoning in Language Models

Explore similar work

Financial Language Models as Applied Artificial Intelligence Systems for News-Based Trading under Market Frictions

Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

Can News Predict the Market? Limits of Zero-Shot Financial NLP and the Role of Explainable AI