acceptodds
Under review as a conference paper at ICLR 2027

FinTriad: A Multimodal, Multi-Agent System with Adaptive Memory for Tri-Horizon Portfolio Construction

Abstract

Professional investing requires integrating multimodal information across heterogeneous decision horizons, while experienced traders accumulate market experience that informs subsequent decisions. Although recent work has increasingly sought to replicate this decision-making process with LLM-based agents, existing systems remain limited in multimodal reasoning, portfolio-level decision-making, and, critically, the continual accumulation of experience from realized outcomes. We introduce FinTriad, a multimodal multi-agent framework with adaptive memory for concurrent tri-horizon portfolio construction. FinTriad integrates broadcast financial media, market-wide candidate selection, specialist analysis, and portfolio management across daily, weekly, and monthly horizons. Its verifiable-reward reflexive learning incorporates realized outcomes into temporally layered memory under a verifiable reward, admitting a correction only when the re-decision that produced it is confirmed against the realized market outcome. Across an 18-month evaluation window spanning multiple market regimes, FinTriad demonstrates competitive portfolio performance and robust adaptation to heterogeneous and evolving market conditions.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multimodal LLM, multi-agent systems, agent memory, verifiable rewards, reflective self-improvement, financial decision-making, portfolio construction

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_r3YsOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FinTriad uses specialist LLM agents to construct daily, weekly, and monthly portfolios from financial broadcasts, charts, and filings. It maintains a history of market events and trading experience. After poor daily outcomes, it generates alternative decisions and admits successful alternatives into a corrective memory. Evaluation consists of a historical U.S. equity backtest, repeated runs, modality and memory ablations, and transaction-cost analysis.

Critical strengths
  1. The paper takes the timing of information seriously. Decisions are frozen before the open, and post-close alternatives affect only subsequent memory rather than being substituted into reported returns. The repeated runs and transaction-cost analysis also make the evidence more useful than a single gross-return backtest.

  2. The memory-removal experiment is a useful attempt to examine the decision process. Holding the current inputs fixed while removing the cited historical entries changes the portfolio manager's decision. Although it is only one case, it provides concrete evidence that accumulated experience can influence later actions across horizons.

Critical weaknesses
  1. [Major] A profitable retrospective alternative does not verify the lesson attached to it. The system diagnoses a failure after observing the outcome and samples alternative decisions against the same realized price path. One alternative clearing a return threshold can reflect luck rather than a correct diagnosis of timing, sizing, or thesis. Moreover, Algorithm 1 still writes information from the best alternative into the general trading memory when verification fails, so the gate does not protect all retained experience. The ablation that admits every diagnosis does not separate useful verification from simply retaining fewer corrections. A comparison that matches memory size while admitting corrections without the return test, or evidence that accepted lessons help on later comparable events, would better support the learning claim. The concern is about the validity of the learned lesson, not the use of future returns in executed trades.

  2. [Major] The baseline design does not isolate better portfolio construction. Giving the baselines an ex post basket may help their asset selection, but it does not equalize their opportunity set: FinTriad can trade additional names, change horizon allocations, and take short positions, while several main baselines are constrained to long-only portfolios. The paper itself attributes part of its resilience during the tariff shock to a short position. FinMem permits shorts, but receives transcript-based inputs and uses a different allocation scheme. The passive return of the excluded names cannot rule out gains from timing or shorting those names. I would like a comparison with common candidates, information, and exposure constraints before attributing the risk and return advantage to the proposed architecture.

Questions for authors

See weaknesses.

Reviewer_4w2MOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents FinTriad, a multimodal multi-agent LLM system for portfolio construction across three concurrent horizons (1-day, 7-day, value/monthly). It has four claimed contributions: (i) direct video ingestion of Bloomberg Brief broadcasts via a multimodal LLM, extracting textual, vocal, and visual cues; (ii) horizon-specific decision cycles driven by a point-in-time data stream; (iii) open-universe candidate selection rather than a predefined shortlist; and (iv) Verifiable-Reward Reflexive Learning (VRRL), which grades each post-close diagnosis against realized prices and writes a correction to memory only if a counterfactual re-decision (up to K=3 samples, never executed) recovers the shortfall (alpha >= 0.3pp vs SPY) that triggered it. The system uses a fixed Gemini 3 Flash backbone (temperature 1.0, knowledge cutoff Jan 2025), a 210-day warm-up, and a 375-day evaluation window (Feb 2025–Jul 2026), reporting CR 112.97%, Sharpe 2.23, MDD 7.46%, outperforming FinRL-A2C/PPO, MVP, 12-1 momentum, RSP, and FinMem baselines, all of which are confined to FinTriad's own 49-stock core basket (90% of exposure-days). Ablations replacing video with audio-only or transcript-only input reduce CR to 22.93% / 13.70%; disabling the VRRL verifier while keeping the diagnosis machinery reduces CR to 40.39% and SR to 1.19.

Critical strengths

S1. VRRL is a genuine conceptual contribution with clean causal evidence. Gating memory writes by a market-priced outcome — rather than the agent grading its own reflection (Reflexion, FinMem) — addresses a documented failure mode of self-graded reflection. The ablation isolates exactly the verifier (diagnosis machinery retained), showing SR drops from 2.23 to 1.19 and MDD rises when unverified lessons are admitted. The CLF case study strengthens this further: with Feb 3 inputs fixed, the PM shorts CLF in 30/30 re-runs with the cited memory entries present, and follows the upstream recommendation in 27/30 runs with those entries removed — evidence that the reversal is a property of accumulated memory, not of the day's evidence.

S2. Experimental hygiene well above the subfield norm. The point-in-time protocol with knowledge-cutoff alignment, the asset-selection control (baselines trading FinTriad's 49-stock core basket, with a complement control showing the excluded 323 names perform like the market), transparent disclosure that FinRL baselines degenerate to near-equal-weight exposure, and cost sensitivity at 5/10 bps together make the comparison architecture credible despite the headline numbers' demanding evidence requirements.

Critical weaknesses

W1. The evidential basis is insufficient for the magnitude of the headline claims.

(a) No statistical significance test (block bootstrap, White Reality Check/Hansen SPA, or Deflated Sharpe Ratio) is reported for any method comparison; the pivotal contrasts — SR 2.23 vs 1.41 (core-basket equal-weight B&H) vs 1.40 (FinMem) — carry no uncertainty quantification.

(b) Ablation results, which carry the paper's RQ3/RQ4 conclusions, are reported as the median of 5 runs with no dispersion.

(c) All results come from one market, one 18-month window (Feb 2025–Jul 2026, an AI-hardware-dominated bull market with a single tariff shock), and one backbone; the Value sleeve alone contributes 84.59% of cumulative return, and the core basket's own passive return is concentrated in five AI-hardware names (MU +806%, INTC +364%, AMD +311%, DELL +300%, FIX +298%). A Sharpe of 2.23 over 18 months is precisely the kind of claim that requires stronger evidence than a single favorable-regime backtest can provide.

W2. Missing the most relevant comparisons, and results are not reproducible. Among LLM-agent baselines only FinMem is evaluated (substantially modified: one agent per ticker, transcript-only news), although the works most directly targeting the claimed gaps — FinAgent (multimodal trading agent), FinCon (multi-agent portfolio management with verbal RL), AlphaAgents (LLM portfolio construction) — are cited but not compared empirically; at least two of the four claimed gaps therefore lack an empirical validation counterpart. Reproducibility is below ICLR expectations for a results-driven paper: closed-source backbone (Gemini 3 Flash), non-redistributable licensed price data and videos, code promised only upon acceptance, and no disclosure of system compute cost. In addition, VRRL's gating constants (α*=0.3pp as both trigger and success criterion; K=3 attempts at temperature 1.0) lack sensitivity analysis and a placebo control, leaving the strictness of the "verifiable" gate unquantified.

Questions for authors
  1. Can the directional conclusions be replicated on a second (ideally open) backbone, even on the 1-day horizon only?
  2. On which data interval were the system's many design constants (position-weight bounds, freshness floor, ±15pp override budget, 40/60 memory-merge ratio, 50-entry cap, etc.) tuned — strictly the warm-up period, or with reference to the evaluation window?
  3. How sensitive are the results to α* ∈ {0, 0.1, 0.3, 0.5} pp and K ∈ {1, 3, 5}? Can you provide a placebo experiment (re-decisions under an uninformative δ_t) estimating the random pass-through rate at temperature 1.0, to calibrate the strictness of the verifier?
Reviewer_zcVxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FinTriad, a multimodal multi-agent framework for portfolio construction across daily, weekly, and monthly investment horizons. The system combines video-based financial news analysis, technical and fundamental specialists, portfolio management, and adaptive memory. Its main methodological component, Verifiable-Reward Reflexive Learning (VRRL), uses realized market outcomes to evaluate counterfactual re-decisions before admitting corrective experience into memory. The authors evaluate the system over 375 trading days following a 210-day warm-up period. They report a cumulative return of 112.97% and a Sharpe ratio of 2.23, together with performance comparisons, transaction-cost sensitivity analyses, and ablations of input modalities and VRRL.

Critical strengths

S1. Integrated Financial Decision Framework. The paper develops a relatively comprehensive system combining multimodal market information, open-universe candidate selection, multiple investment horizons, and adaptive memory (Sections 3.1–3.4). The explicit point-in-time protocol in Appendix B.7 is particularly valuable for evaluating financial agents.

S2. Meaningful Empirical Analysis. The evaluation includes repeated runs, transaction-cost sensitivity, modality ablations, and VRRL ablations (Tables 1–4). These experiments provide useful evidence beyond a single reported backtest. The explicit separation between executed trading returns and post-close counterfactual learning is also a positive design choice.

Critical weaknesses

W1. Insufficient Identification of the Claimed Portfolio-Construction Contributions (Fundamental). The comparison with strategy baselines does not fully isolate the contributions of open-universe selection, multi-horizon coordination, and portfolio construction. In Section 4.1 and Appendix D, the 49-stock baseline universe is selected retrospectively using FinTriad's exposure over the entire evaluation period, whereas FinTriad itself trades 372 stocks. Although this choice can favor the baselines by granting them access to FinTriad's selected assets, it does not establish a common ex-ante opportunity set or disentangle the proposed components. Furthermore, Appendix C attributes 84.59 of the 112.97 percentage points of cumulative return to the monthly value horizon. The independent contribution of tri-horizon coordination therefore remains insufficiently established. Stronger component-level comparisons under common ex-ante investment universes are needed.

W2. Limited Evidence That VRRL Produces Generalizable Learning (Fundamental). The central VRRL mechanism selects counterfactual corrections based on their performance against the same realized prices used to diagnose the original decision (Section 3.4, Algorithm 1). Although the selected corrections affect only subsequent trading days, this retrospective selection can still overfit individual outcomes. Moreover, unsuccessful corrections may still influence the broader trading memory (Appendix B.6). The reported ablation compares verified reflection against admitting unverified reflections, rather than isolating the value of outcome-based verification from simpler memory updates or the reflection mechanism itself. The method is evaluated with one backbone over one market period, leaving the generalizability of the claimed learning mechanism uncertain. The 10-bps cost results, where cumulative return falls from 112.97% to 62.14% and below several strategy baselines, further limit the robustness of the economic conclusions.

Questions for authors
  1. Contribution identification: Are there existing results for a value-only configuration and a multi-horizon configuration with fixed or random asset selection, evaluated under the same portfolio constraints? These would help establish whether the claimed benefits arise from cross-horizon coordination rather than the dominant value sleeve or asset-selection decisions.

  2. VRRL generalization: How frequently do verified corrections improve subsequent unseen decisions, rather than merely matching the realized outcome used for their selection? Can the authors clarify the separate effects of verified corrections and the broader memory updates derived from unsuccessful attempts?

  3. Economic robustness: Given the sharp deterioration at 10-bps transaction costs, what assumptions justify the reported trading costs and execution conditions? Are turnover and net performance available separately for each investment horizon?

Reviewer_zYX8OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents FinTriad, a multimodal, multi-agent framework for portfolio construction across daily, weekly, and monthly horizons. It combines financial news videos, market data, and SEC filings for stock selection and trading decisions. Its key component, Verifiable-Reward Reflexive Learning (VRRL), evaluates counterfactual trading decisions against realized market returns and selectively stores corrections in memory without updating the underlying LLM. The authors evaluate FinTriad over 375 trading days against several baselines, reporting a cumulative return of 112.97% and a Sharpe ratio of 2.23. Ablation studies assess the effects of multimodal inputs and VRRL, while a case study illustrates how stored corrections influence decisions across investment horizons.

Critical strengths

FinTriad provides an end-to-end framework that integrates multimodal financial information, stock selection, and portfolio management across three investment horizons. Its point-in-time data handling and shared memory demonstrate how LLM agents can be combined to support a complex financial decision-making process.

Critical weaknesses
  1. FinTriad combines multimodal analysis, multi-horizon trading, stock selection, and adaptive memory, but its contribution beyond system integration remains unclear. Although the paper demonstrates cross-horizon memory reuse, it does not directly show whether these components provide additional benefits through their interaction.
  2. VRRL verifies counterfactual decisions using portfolio-level returns, but stores corrections for individual positions. A successful portfolio does not necessarily mean that each correction is valid. While the ablation shows that verification improves performance over accepting all corrections, it remains unclear whether the proposed criterion is better than simpler memory selection strategies.
Questions for authors
  1. Beyond the CLF case study, is there evidence that sharing memory across horizons improves overall performance compared with keeping their memories separate?
  2. VRRL uses portfolio-level returns to verify position-specific corrections. How does the method ensure that each stored correction is actually valid?
  3. Could VRRL's improvement come from storing fewer corrections rather than the verification criterion itself? Is there a comparison with simpler selection rules under a similar memory budget?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.