technical indicator engineering

Existing LLM-driven multi-agent financial trading systems suffer from insufficient agent coordination, absence of structured self-reflection mechanisms, and scarcity of high-quality domain-specific training data. To address these challenges, this paper proposes a self-reflective, multi-agent trading framework with automated domain data synthesis capabilities. Specifically, we design a collaborative decision-making mechanism based on signal fusion and dynamic risk management; introduce self-reflection prompting and trading-style adaptation modules to enable cross-scenario experience transfer; and construct an end-to-end domain data synthesis pipeline that generates high-fidelity, labeled data incorporating market states and agent decisions. Backtesting on five real-world stock datasets demonstrates that our system significantly outperforms rule-based, machine learning, reinforcement learning, and state-of-the-art LLM-based baselines—validating its superiority in decision quality, market adaptability, and agent coordination.

Existing LLM-based financial agents predominantly rely on sentiment or trend analysis, lacking principled risk modeling mechanisms, which compromises decision-making accuracy. Method: We propose the first intelligent agent framework integrating large language models (LLMs) with agent-based model discovery (ABMD), enabling iterative inference of stochastic differential equations (SDEs) governing financial time series. This yields interpretable, automated market risk modeling and subsequent strategy generation. The approach synergizes LLM reasoning, tool-augmented execution, symbolic SDE discovery, and market simulation to achieve end-to-end closed-loop operation in both real and synthetic environments. Contribution/Results: Empirical evaluation demonstrates that our strategy significantly improves the Sharpe ratio across multiple stocks in backtesting, consistently outperforming baseline LLM-driven trading agents. To our knowledge, this is the first work unifying LLM-powered financial dynamical modeling with risk-aware decision-making—establishing a foundation for interpretable, physics-informed AI trading.

This paper addresses the challenge of jointly modeling temporal technical indicators and static fundamental information—tasks poorly handled by single-model approaches. We propose a hybrid LSTM–Random Forest forecasting framework: an LSTM module captures deep sequential patterns from price time series, while a Random Forest integrates technical indicators (e.g., MACD, RSI) with macroeconomic and firm-level fundamentals; crucially, it incorporates a feature-importance-driven technical indicator selection mechanism. Evaluated on 10-day return prediction for international public companies, our method significantly outperforms baseline models—including standard LSTM, Random Forest, and XGBoost—in both predictive accuracy (p < 0.01) and out-of-sample Sharpe ratio. Results demonstrate that heterogeneous data fusion yields substantial, statistically robust gains in quantitative trading performance. The framework offers a novel, interpretable, and robust paradigm for intelligent trading powered by multi-source financial data.

To address the poor stability and weak adaptability of deep learning models in quantitative investing, this paper proposes an LLM-driven multi-agent collaborative framework for automated discovery and dynamic ensemble optimization of multimodal (numerical, textual, and chart-based) alpha factors. The method innovatively integrates large language models (LLMs), multi-agent systems, and a dynamic weight gating mechanism to establish a market-state-aware, adaptive strategy generation paradigm. Empirically evaluated on the Chinese A-share market, the framework significantly outperforms state-of-the-art baselines: it achieves a 23.6% improvement in Sharpe ratio and a 31.2% reduction in maximum drawdown, effectively balancing return enhancement and risk control. By enabling interpretable, robust, and adaptive decision-making, the proposed framework establishes a novel paradigm for AI-powered quantitative investment.

Traditional technical analysis suffers from limited predictive accuracy in financial markets characterized by nonlinearity, high noise, and frequent exogenous shocks. To address this, we propose ElliottAgents—a novel interpretable AI trading system that is the first to deeply integrate Elliott Wave Theory into a large language model (LLM)-based multi-agent framework, synergizing retrieval-augmented generation (RAG), deep reinforcement learning (DRL), and multi-agent coordination. Our system enables automated wave-pattern recognition, cross-temporal trend inference, and decision traceability, thereby bridging interpretability and adaptability in AI-driven trading. Empirical evaluation on historical U.S. equity data demonstrates that ElliottAgents significantly outperforms baseline methods: waveform identification accuracy and trend prediction stability are markedly improved, with signal accuracy increasing by 23.6% and strategy Sharpe ratio rising by 31.4%.

This work addresses the fragmentation of existing AI systems in finance, which typically operate in isolation and struggle to jointly handle tasks such as robo-advising, high-frequency trading, dynamic recommendation, strategic banking interactions, and cross-modal sentiment analysis. To overcome this limitation, we propose the first unified multimodal intelligent finance framework that integrates proximal policy optimization (PPO)-based reinforcement learning, time-series forecasting, in-context learning, Nash equilibrium computation from game theory, and cross-modal embeddings. We further design an ensemble optimization method with convergence guarantees. Extensive experiments demonstrate that our framework achieves significant improvements: a 23.7% gain in portfolio optimization performance, a 31.2% reduction in high-frequency trading error, an 18.9% increase in investment recommendation accuracy, a 27.4% acceleration in game-theoretic convergence speed, and a 15.6% boost in sentiment analysis accuracy—collectively unlocking the synergistic potential of multifaceted AI techniques in finance.

This study addresses the limitations of existing financial sentiment analysis methods, which rely on static human annotations and struggle to adapt to dynamic market conditions. To overcome this, the authors propose the first market-aligned reinforcement learning framework that replaces manual labels with real market feedback, enabling an end-to-end optimization pipeline for financial large language models. The approach incorporates market-aware data filtering, a discrete asymmetric trading reward mechanism, and dynamic retraining at arbitrary time points to adaptively refine sentiment signals. Empirical results demonstrate that the proposed method significantly outperforms state-of-the-art baselines in terms of profitability, risk-adjusted returns, and sentiment signal quality, achieving a 220% improvement in cumulative trading returns over the strongest baseline.

Current research on LLM-driven trading systems is widely hindered by the absence of comparable evaluation protocols, well-defined execution semantics, and reproducibility, significantly impeding progress in the field. This work proposes an Architecture–Capability–Adaptation analytical framework to systematically review 77 studies, employing protocol encoding snapshots, empirical categorization, R0–R3 reproducibility ratings, and transaction semantics auditing to construct an evidence ledger and a standardized reporting checklist. The analysis reveals that only a minority of studies satisfy closed-loop evaluation criteria, while most lack temporally consistent data splits, explicit transaction cost modeling, or proper handling of survivorship bias. These findings underscore the urgent need for standardized evaluation practices and provide a reproducible benchmark and methodological guidance for future research.

This study addresses the challenge of accurately capturing short-term stock trend dynamics in non-stationary and nonlinear markets, where traditional trend-following strategies often falter. The authors propose an LSTM-based predictive framework that incorporates time series differencing to reduce both bias and variance in forecasts, specifically targeting the next-day trend differential (Δt) for the top 30 constituents of the S&P 500. Evaluated across multiple market cycles from 2005 to 2025, the proposed method is systematically benchmarked against OLS, Ridge, Lasso, and LightGBM models. Empirical results demonstrate that the LSTM-driven approach yields significantly higher economic returns in portfolio simulations, consistently outperforming all baseline models in terms of overall profit-and-loss (PNL) metrics, thereby confirming its effectiveness and robustness in dynamic market environments.

This work addresses the inefficiencies in decentralized prediction markets—specifically, inadequate information aggregation, mispricing, and transaction latency—that collectively erode arbitrage opportunities. To tackle these challenges, the paper introduces the first prediction framework based on a multi-agent large language model (LLM) architecture, integrating 50 diverse LLM agents. The system employs confidence-weighted Bayesian fusion to combine market-implied probabilities with crowd consensus, executes trades under risk constraints using a quarter-Kelly criterion, and identifies cross-market pricing inefficiencies via Kullback–Leibler and Jensen–Shannon divergences. Furthermore, it leverages a log-normal model to infer centralized exchange prices, enabling latency arbitrage within human reaction-time windows. Empirical results on Polymarket demonstrate substantially improved probability calibration, with Brier scores and log-loss approaching those of human superforecasters and significantly outperforming single-model baselines.