FORESIGHT-9: Prospective and Process-Aware Evaluation of Adaptive Trading Agents

Xiangxin Luo; Chengtian Hong; Haohua Li; Yongyi Xie

cs.AIAug 29, 2026

FORESIGHT-9: Prospective and Process-Aware Evaluation of Adaptive Trading Agents

Authors: Xiangxin Luo, Chengtian Hong, Haohua Li, Yongyi Xie

Organizations: University of Science and Technology of China · University of Hong Kong · Shanghai Jiao Tong University

Abstract

Retrospective backtests provide a limited test of adaptive trading agents: they cannot rule out historical contamination, expose sensitivity to a single realized market path, or reveal internal degeneration during long-horizon adaptation. We introduce FORESIGHT-9, a prospective and process-aware benchmark built from nine auditable counterfactual stress worldlines branching from a common July 2026 information boundary. Each worldline specifies staged macro-financial events and joint multi-asset anchors; a deterministic generator realizes the trajectories, while observations are disclosed according to in-world time. A common contract standardizes observations and execution while preserving each agent's native adaptation loop. We evaluate two adaptive trading-agent frameworks with two foundation-model backbones across 36 long-horizon runs. Agent rankings vary substantially across worldlines and backbones, and a fixed equal-weight policy outperforms 31 of 36 runs. Process telemetry exposes failures that terminal returns conceal: in one high-return run, the live factor library collapsed while executed holdings converged to the equal-weight fallback, even though decision records continued to report an active factor ensemble. FORESIGHT-9 therefore evaluates not only portfolio outcomes, but whether adaptive agent state and execution remain coherent across alternative futures. We release the worldlines, trajectories, audit traces, and regeneration scripts.

Figures & tables

Figure 1: FORESIGHT-9 overview. The shared pre-boundary warmup produces the learned research and policy state without scored portfolio evaluation. At the information boundary, the frozen state supports both the deliberately in-sample Retrospective Replay (RR) diagnostic and prospective online adaptation across nine worldlines. Outcomes and factor/trading records are audited separately, while the contract fixes temporal progression, portfolio constraints, costs, and the account-write interface. In the figure, 2026-07-16 is the shared observed execution/market anchor and generator handoff; scenario-specific synthetic rows begin on 2026-07-17.

WL	Worldline crisis	Main transmission
WL1	Geopolitical	conflict $\to$ sanctions $\to$ trade fragmentation
WL2	Technology fragmentation	internet splintering $\to$ bipolar digital economy
WL3	Nuclear escalation	nuclear test $\to$ conventional conflict $\to$ new nuclear order
WL4	Sovereign debt / liquidity	Treasury stress $\to$ Minsky moment $\to$ stagflation
WL5	Currency disorder	JPY disorderly devaluation $\to$ Asian currency fragmentation
WL6	Inflation regime	inflation re-ignition $\to$ credibility collapse $\to$ stagflation

Table 1: The nine worldlines at a glance. Each worldline is a sequence of 4–5 stages; the full stage end dates and narratives are in Table 8 (Appendix C ). All WLs end 2035-12-31.

Configuration	RR ann. return	RR rank	Prospective median rank	WLs ranked 1st
FM–DS	21.64%	1	4	1/9
AC–terra	18.81%	2	3	0/9
FM–terra	17.19%	3	1	5/9
AC–DS	16.09%	4	2	3/9

Table 2: Retrospective-to-prospective rank transfer. RR ranks configurations by terminal cumulative return within the realized-history replay (the annualized return is shown only as a historical-performance anchor); the prospective columns summarize agent-only ranks across WL1–WL9. Full RR portfolio metrics are reported in Appendix E .

Figure 2: Retrospective-to-prospective rank transfer. Each cell shows the agent-only rank (1 = best) under Native-Policy RR or within a prospective worldline. The realized-history ordering transfers inconsistently across worldlines (mean Spearman $\bar{\rho}=-0.42$ , median $\rho=-0.80$ ); the RR leader FM–DS remains first on only one of nine worldlines.

Figure 3: Cross-worldline mean normalized prospective NAV. The four agent–backbone configurations are compared with the periodically rebalanced equal-weight reference (EW-15). At each in-world date, each run is normalized by the common boundary capital and averaged with equal descriptive weight across WL1–WL9; the curve is a cross-scenario summary, not a probability-weighted expected future path. Deterministic baselines are reported separately in Appendix A .

Appendix figures & tables33 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Extended cross-worldline mean normalized prospective NAV. The four adaptive configurations and three deterministic references (EW-15, buy-and-hold EW, and inverse-volatility 60d) are shown on a common boundary-capital normalization. This aggregate view averages the normalized daily NAV pointwise across the nine fixed worldlines with equal descriptive weight; it is a cross-scenario summary rather than a probability-weighted or realizable future path. The worldline-level baseline heterogeneity is shown separately below.

Figure 5: Deterministic baseline comparison across prospective worldlines. The figure isolates EW-15, buy-and-hold EW, and inverse-volatility (60d); cumulative returns are computed under the same shared information boundary, 2026-07-16 observed anchor, worldline panels, and fee convention as the agent runs.

Figure 6: AC–terra across the nine worldlines (nine solid runs; colorblind-validated palette; right-edge labels mark the terminal worldline). Conventions are identical in Figures 7 – 18 .

Figure 7: AC–DS across the nine worldlines.

Figure 8: FM–terra across the nine worldlines.

Figure 9: FM–DS across the nine worldlines. The framework that leads under terra (Figure 8 ) trails under DS: the per-worldline ranking is different across the two evaluated backbones.

Figure 10: WL1: four runs vs. the equal-weight baseline. AC–terra (blue), AC–DS (pink), FM–terra (amber), FM–DS (green), EW-15 (gray); labels give terminal returns.

Figure 11: WL2: four runs vs. the equal-weight baseline. AC–terra (blue), AC–DS (pink), FM–terra (amber), FM–DS (green), EW-15 (gray); labels give terminal returns.

Figure 12: WL3: four runs vs. the equal-weight baseline. AC–terra (blue), AC–DS (pink), FM–terra (amber), FM–DS (green), EW-15 (gray); labels give terminal returns.

Figure 13: WL4: four runs vs. the equal-weight baseline. AC–terra (blue), AC–DS (pink), FM–terra (amber), FM–DS (green), EW-15 (gray); labels give terminal returns.

Figure 14: WL5: four runs vs. the equal-weight baseline. AC–terra (blue), AC–DS (pink), FM–terra (amber), FM–DS (green), EW-15 (gray); labels give terminal returns.

Figure 15: WL6: four runs vs. the equal-weight baseline. AC–terra (blue), AC–DS (pink), FM–terra (amber), FM–DS (green), EW-15 (gray); labels give terminal returns.

Figure 16: WL7: four runs vs. the equal-weight baseline. AC–terra (blue), AC–DS (pink), FM–terra (amber), FM–DS (green), EW-15 (gray); labels give terminal returns.

Figure 17: WL8: four runs vs. the equal-weight baseline. AC–terra (blue), AC–DS (pink), FM–terra (amber), FM–DS (green), EW-15 (gray); labels give terminal returns.

Figure 18: WL9: four runs vs. the equal-weight baseline. AC–terra (blue), AC–DS (pink), FM–terra (amber), FM–DS (green), EW-15 (gray); labels give terminal returns.

Worldline	AC-terra	AC-DS	FM-terra	FM-DS	EW-15
WL1	$+34.9\%$	$+57.5\%$	$+53.4\%$	$+34.9\%$	$+66.6\%$
WL2	$+43.4\%$	$+49.9\%$	$+87.8\%$	$+41.4\%$	$+77.0\%$
WL3	$+38.6\%$	$+62.4\%$	$+96.8\%$	$+13.8\%$	$+54.7\%$
WL4	$+5.8\%$	$+25.1\%$	$+8.8\%$	$+20.3\%$	$+25.2\%$
WL5	$+39.4\%$	$+59.4\%$	$+138.7\%$	$+8.4\%$	$+73.9\%$
WL6	$+72.2\%$	$+67.0\%$	$+77.1\%$	$+85.7\%$	$+103.9\%$

Appendix

Table 3: Full prospective outcome matrix. Cumulative total return per worldline for the four agent–backbone combinations and the equal-weight baseline. Within each worldline, all four configurations used the same byte-identical market panel, portfolio constraints, and fee rule.

	AC-terra	AC-DS	FM-terra	FM-DS	EW-15
WL1	0.41	0.52	0.37	0.27	0.59
WL2	0.53	0.41	0.61	0.35	0.78
WL3	0.46	0.58	0.67	0.14	0.60
WL4	0.10	0.39	0.10	0.23	0.35
WL5	0.50	0.55	0.81	0.09	0.72
WL6	0.67	0.50	0.54	0.64	0.92

Appendix

Table 4: Sharpe ratio per worldline (daily NAV, annualized,

r_{f}=0

	AC-terra	AC-DS	FM-terra	FM-DS	EW-15
WL1	-28.4%	-29.0%	-35.4%	-34.1%	-30.9%
WL2	-16.9%	-19.1%	-27.0%	-27.8%	-14.6%
WL3	-14.2%	-25.7%	-27.5%	-28.0%	-26.7%
WL4	-26.2%	-18.6%	-25.0%	-32.7%	-22.9%
WL5	-10.3%	-13.0%	-14.7%	-26.9%	-12.5%
WL6	-16.5%	-18.1%	-20.7%	-18.5%	-14.9%

Appendix

Table 5: Maximum drawdown per worldline (daily NAV).

Strategy	Mean return	Median return	Median Sharpe	WLs $>$ EW-15	Paired wins vs. agents (/36)
EW-15	$+70.2\%$	$+73.9\%$	0.70	—	31/36
Buy-and-hold EW	$+22.0\%$	$+13.5\%$	0.13	0/9	3/36
Inverse-volatility (60d)	$+70.4\%$	$+74.9\%$	0.69	4/9	31/36

Appendix

Table 6: Deterministic reference comparison. Returns are cumulative prospective returns summarized across WL1–WL9; Sharpe is the median of the per-worldline daily-NAV Sharpe ratios. Agent comparator returns use the fixed initial-capital denominator

V_{0}=1{,}000{,}000

(terminal NAV divided by

V_{0}

, minus one), matching the paper convention.

Field	Warmup (real data)	Forward (synthetic)
close	observed	generator output (Eq. 1 , re-anchored)
open	observed	previous close
high / low	observed	$\max/\min(\text{open},\text{close})\times(1\pm\mathrm{rng}/2)$
volume	observed	per-asset warmup median (constant)
amount	observed	close $\times$ volume
vwap	amount $/$ volume	equals close

Appendix

Table 7: Observation schema, warmup vs. forward. rng is the per-asset median of

(\text{high}-\text{low})/\text{close}

over the last 60 pre-boundary days; high/low are undefined for the seven zero-amplitude series listed above, and volume is zero (hence vwap undefined) for the yield, FX, and VIX series.

WL	Stage end dates	Stage narratives (in order)
WL1	27-12, 28-05, 29-12, 31-12, 35-12	geopolitical escalation: red-line warnings $\to$ 72-hour blitz $\to$ sanctions & global restructuring $\to$ new equilibrium $\to$ new normal
WL2	27-12, 29-12, 31-12, 35-12	internet splintering $\to$ fintech-system splintering $\to$ bipolar consolidation $\to$ bipolar digital economy
WL3	27-12, 28-06, 29-12, 31-12, 35-12	nuclear test escalation $\to$ conventional conflict $\to$ deterrence failure & restructuring $\to$ non-proliferation rebuild $\to$ new nuclear order
WL4	27-12, 29-12, 31-12, 35-12	cracks spreading $\to$ Minsky moment $\to$ stagflation & de-dollarization $\to$ early institutional rebuild
WL5	27-12, 29-12, 31-12, 35-12	cracks emerge $\to$ JPY disorderly-devaluation spiral $\to$ Asian currency fragmentation $\to$ Japan’s “new normal”
WL6	27-12, 29-12, 31-12, 35-12	inflation re-ignition $\to$ central-bank credibility collapse $\to$ entrenched stagflation $\to$ new monetary order

Appendix

Table 8: The nine worldlines: full stage table. Stage terminal dates (YY-MM) and narratives, represented in the machine-readable JSON specification. WL9’s third stage carries no explicit numeric date ( † ; its narrative lies between 2029-12 and 2033-12). All WLs end 2035-12-31.

Agent column	Window hit rate	Mean excess / window
	(%, descriptive interval)	(bps, descriptive interval)
AC-terra	45.0 $[43.5,\,46.6]$	$-8.2$ $[-9.0,\,-7.2]$
AC-DS	47.4 $[45.4,\,49.5]$	$-5.7$ $[-9.9,\,-2.1]$
FM-terra	49.9 $[49.0,\,50.8]$	$+2.3$ $[-2.0,\,+6.9]$
FM-DS	48.7 $[46.5,\,51.0]$	$-10.3$ $[-15.0,\,-6.0]$
Pooled	47.8 $[46.7,\,48.9]$	$-5.5$ $[-7.0,\,-4.0]$

Appendix

Table 9: Per-window comparison against EW-15 (245 non-overlapping ten-trading-day windows per run; 8,820 paired comparisons). Hit rate is the share of windows the agent beats EW-15; excess is the agent’s window return minus EW-15’s, in basis points. Intervals are descriptive worldline-clustered bootstrap summaries.

Configuration	Boundary reference	$\rho$
AC–terra	archived native first decision	1
AC–DS	archived native first decision	1
FM–terra	archived native first decision	1
FM–DS	nine archived WL references	1

Appendix

Table 10: Native boundary reconstruction golden tests. Exact comparisons use the fixed asset set and order; source and bundle hashes are recorded in the artifact manifest.

Configuration	Cum.	Ann.	Vol.	Sharpe	MDD	Proposals	Executed	Proposal turnover	Cost (bps of $V_{0}$ )	Agent rank
FM–DS	275.88%	21.64%	17.25%	1.223	-23.56%	157	156	68.40	499.0	1
AC–terra	220.51%	18.81%	12.01%	1.496	-12.43%	157	8	22.42	1.8	2
FM–terra	192.15%	17.19%	15.45%	1.105	-32.04%	157	156	71.23	389.0	3
AC–DS	174.03%	16.09%	10.52%	1.471	-11.89%	157	145	15.65	77.4	4
EW–15	227.47%	19.19%	14.41%	1.292	-22.38%	157	157	3.60	19.0	—

Appendix

Table 11: Full RR metrics on the common realized-history window. Cum. is terminal return, Ann. is annualized return, Vol. is annualized volatility, and MDD is maximum drawdown. Proposal turnover sums one-way proposed migration, including gated-out proposals; executed counts and costs refer only to accepted rebalances. Cost is in basis points of initial benchmark capital. RR and prospective cumulative returns have different horizons and are not compared by raw magnitude.

WL	Spearman $\rho$	Preserved pairs	FM–DS WL rank
1	-0.8	1/6	3
2	-0.8	1/6	4
3	-0.8	1/6	4
4	-0.4	2/6	2
5	-0.8	1/6	4
6	+0.8	5/6	1

Appendix

Table 12: Prospective rank transfer for the four agent configurations. Rankings are defined by terminal cumulative return within each environment. Statistics are descriptive: no p-values or significance tests are applied.

Configuration	RR excess sign	WLs $>$ EW	Median WL excess	Median WL rank
AC–terra	negative	0/9	-2.14 pp	3
AC–DS	negative	1/9	-0.94 pp	2
FM–terra	negative	4/9	-0.41 pp	1
FM–DS	positive (+2.46 pp)	0/9	-2.41 pp	4

Appendix

Table 13: Apparent excess return relative to each environment’s EW-15 reference. The RR and prospective environments use different horizons, so only within-environment excess and signs are interpreted.

Panel of Figure 19 — Figure 19: The admission funnel. Gate values are stipulated and deliberately permissive; selectivity comes from the staged tightening. Gate 3 counts the candidate’s pairwise Spearman conflicts ( $\rho>0.5$ ) with incumbents: $k\geq 2$ rejects the candidate; $k=1$ triggers a quality duel on $q=|\mathrm{IC}|\cdot|\mathrm{ICIR}|$ with the conflicting incumbent, the winner retained; $k=0$ admits it, with the capacity-30 trim evicting the lowest- $q$ member. Gray rates are measured on the warmup candidate pool and on online screen-passers.

	AC–terra	AC–DS	FM–terra	FM–DS
Warmup qualified library (enters every WL)	6	4	21	8
Online-admitted terminal members listed per WL ∗	270 (30 per WL; all distinct)	99 (3–20 per WL; 94 distinct)	11 (on 7 WLs)	0
Terminal live-library size across WLs	30 (at capacity)	0–22	21–24	8 (warmup only)
Mining-screen candidates ( $\Sigma\geq 3$ ), never admitted	—	—	—	109 ( $\Sigma$ 824)

Appendix

Table 14: Factor lifecycle in numbers. ∗ AC–DS’s WL8 is the citation-listed exception: its library was evicted to zero before the horizon (§ 5.4 ). FM–DS admitted nothing online; its 109 unadmitted candidates accumulated

\Sigma=824

screenings across the nine mining streams. The reported inventory summaries regenerate from the released factor-lifecycle and factor-library audit logs.

Experiment	Factor	IC / ICIR	Formula
AC–terra	peer_median_leadlag_5d	+0.056 / 0.17	median({return_5d[j] for j != i})
AC–terra	miner_3_clv_1d	+0.068 / 0.19	-(2 * (close - low) / (high - low) - 1)
AC–DS	mom_10d_skip5	+0.041 / 0.12	close.shift(5) / close.shift(15) - 1.0
AC–DS	vix_beta_cond_60x20	-0.038 / -0.09	-beta(asset_ret, VIX_ret, 60) * (VIX/VIX.shift(20) - 1.0)
FM–terra	return_rank_mean_reversion	+0.031 / 0.09	Neg(CsRank(TsRank(Sum($returns, 5), 30)))
FM–terra	intraday_vwap_pressure	+0.028 / 0.09	CsZScore(Mul(Div(Sub( $close,$ vwap), Add(Abs( $vwap), 0.001)), Div(Sub($ vwap, $open), Add(Abs($ open), 0.001))))

Appendix

Table 15: Representative warmup factors (IC/ICIR from the warmup evaluation; each library seeds all nine worldlines of its experiment). The complete lists of 6/4/21/8 factors are in the released manifest.

Stage / stream	Factor	Evidence	Formula
AC–terra, WL1	miner_1_20281116_defensive_relative_lead_20d	cited 29-04-19 $\sim$ 35-12-21 (144 dec.)	($asset_return_20d - mean(XAU, US10Y, CN10Y 20d)) * (0.25 + 0.75*I(breadth_20d < 0.40)), lagged
AC–terra, WL1	macro_stress_resilience_20d	cited 29-12-13 $\sim$ 35-12-21 (100 dec.)	shift1((r20 - median_cs(r20)) / (downside_vol20 + 0.02) * (1 + 0.35clip((VIX-median60(VIX))/std60(VIX),-1,2)tanh(5*(r20-median_cs(r20)))))
AC–DS, WL2	spx_corr60	cited 26-08-13 $\sim$ 35-12-06 (198 dec.)	close.pct_change().rolling(60).corr(close[’SPX’].pct_change())
AC–DS, WL3	cn10y_beta_60	cited 26-07-30 $\sim$ 35-12-20 (227 dec.)	rolling_beta(asset_daily_ret, cn10y_yield_daily_change, 60)
FM–terra, WL4	range_adjusted_return_rank	admitted 2026-08-01	CsRank(Div(TsRank( $returns, 13), Add(TsRank(Div(Sub($ high, $low), Add(Abs($ close), 0.001)), 13), 0.001)))
FM–terra, WL5	volume_acceleration_pressure	admitted 2026-08-03	CsZScore(Delta(EMA(Log(Add($volume, 1)), 6), 14))

Appendix

Table 16: Representative online admissions and mining candidates. AC spans show first–last decision citing the factor (

n

decisions); FM–terra gives the admitting WL and timestamp; mining rows give the total screen count

\Sigma

across the nine FM–DS mining streams (all failed the correlation gate; none was ever admitted).

WL	dec	reb	cost	lib $\to$ ens	dec	reb	cost	lib $\to$ ens
	AC-terra				AC-DS
1	43	205	30.9	30 $\to$ 8	35	212	100.3	14 $\to$ 6
2	121	159	33.8	30 $\to$ 9	47	200	102.8	22 $\to$ 5
3	294	134	31.1	30 $\to$ 7	19	228	71.8	22 $\to$ 10
4	98	45	16.1	30 $\to$ 5	64	187	45.2	3 $\to$ 2
5	76	159	37.8	30 $\to$ 7	235	12	3.7	12 $\to$ 0

Appendix

Table 17: AC process registers. dec = declined proposals: each cycle’s proposed portfolio shift is ruled on by comparing its expected gross edge against friction (migration) cost, and every dec entry is a rejection by that ruling (tagged gross_edge_not_above_migration_cost , unchanged target, or missing forecast); each entry carries the proposed weights, forecasts, and factor ids. reb = executed rebalances: the initial build plus every proposal the gate accepted, each with its transferred notional and realized cost. The two registers partition the proposal stream under the shared gate (see below); cost = cumulative transaction cost in bps of initial capital over 9.5 years (3 bps on migrated notional, 0 on the initial build); lib

\to

ens = terminal live-library size

\to

ensemble-selected factors. FM terminal libraries hold 21–24 factors (terra) and exactly the 8 warmup factors (DS) in every WL; the staged inventory—warmup libraries, online admissions, and the DS mining supplement—is in Appendix F .

WL	Theme	Rigor (/5)	10-yr plaus. (%)	Severity (/10)
1	geopolitical escalation	3	10–15	9
2	tech “silicon curtain”	4	40–50	7.5
3	nuclear-proliferation chain	3	8–12	8.5
4	Treasury “Minsky moment”	3.5	18–25	10
5	JPY devaluation spiral	3	15–22	7
6	stagflation & CB collapse	3	10–15	8

Appendix

Table 18: Post-debate review scores per worldline , documenting the scenario workflow: logical rigor (5-point scale), agent-elicited 10-year plausibility range, and shock severity (10-point scale).

Model	Reasoning mode	Reasoning effort	Temperature	Top-p
gpt-5.6-terra	standard	medium	default	default
deepseek-v4-flash	standard	high	default	default

Appendix

Table 19: LLM configuration for all runs.

Long-Horizon Agent Evaluation

Long-Horizon Agent Evaluation measures whether AI agents can plan, act, remember, and stay reliable across extended tasks and changing environments.

See all 355 papers

TraderBench: How Robust Are AI Agents in Adversarial Capital Markets?
Feb 27, 2026Xiaochuang Yuan, Hui Xu, Silvia Xu +2AI Agent Evaluation AI Agent Benchmarks
UC Santa Cruz
FutureSim: Replaying World Events to Evaluate Adaptive Agents
May 14, 2026Shashwat Goel, Nikhil Chandak, Arvindh Arun +5Continual Test-Time Adaptation Forecasting Benchmarks
1ELLIS Institute Tübingen · 2Max Planck Institute for Intelligent Systems · 4Tübingen AI Center +3
Agentic Empirical Asset Pricing: Methodological Foundations
Sep 1, 2026Yingjian Pan, Xiaowei Ding, Kay GieseckeQuantitative Finance AI Agent Evaluation
Advanced Financial Technologies Laboratory · Stanford University · School of Information Management +1

Feb 27, 2026J/K move · Enter open · S save

TraderBench: How Robust Are AI Agents in Adversarial Capital Markets?

Xiaochuang Yuan, Hui Xu, Silvia Xu, Cui Zou+1

UC Santa Cruz

AI Agent Evaluation AI Agent Benchmarks Adversarial Robustness Algorithmic Trading Quantitative Finance

Open paper AskPDF arXiv

Evaluating AI agents in finance faces two key challenges: static benchmarks require costly expert annotation yet miss the dynamic decision-making central to real-world trading, while LLM-based judges introduce uncontrolled variance on domain-specific tasks. We introduce TraderBench, a benchmark that addresses both issues. It combines expert-verified static tasks (knowledge retrieval, analytical reasoning) with adversarial trading simulations scored purely on realized performance-Sharpe ratio, returns, and drawdown-eliminating judge variance entirely. The framework features two novel tracks: crypto trading with four progressive market-manipulation transforms, and options derivatives scoring across P&L accuracy, Greeks, and risk management. Trading scenarios can be refreshed with new market data to prevent benchmark contamination. Evaluating 13 models (8B open-source to frontier) on ~50 tasks, we find: (1) 8 of 13 models score ~33 on crypto with <1-point variation across adversarial conditions, exposing fixed non-adaptive strategies; (2) extended thinking helps retrieval (+26 points) but has zero impact on trading (+0.3 crypto, -0.1 options). These findings reveal that current agents lack genuine market adaptation, underscoring the need for performance-grounded evaluation in finance.

Abstract

Long-Horizon Agent Evaluation

Explore similar work

TraderBench: How Robust Are AI Agents in Adversarial Capital Markets?

FutureSim: Replaying World Events to Evaluate Adaptive Agents

Agentic Empirical Asset Pricing: Methodological Foundations