Literature Review

The intellectual foundations of adaptive, regime-aware, risk-governed market decision-making — and where this proposal sits among them

A thematic, critical review of the seven literatures the proposal integrates — market efficiency, regime detection, volatility, reinforcement learning, NLP and LLMs in finance, multi-agent systems, and backtest methodology — positioned against the 2024–2026 wave of multi-agent LLM trading systems whose own critics have begun to document the gap this proposal sets out to fill.

This review organises the field into the seven literatures the proposal integrates, reads each critically rather than cataloguing it, and locates the proposal against a body of work that barely existed when the idea was first sketched: the 2024–2026 explosion of multi-agent LLM trading systems. Its central argument is that the contribution is not another such system — those now exist and are proliferating — but a control-first, regime-aware, risk-governed and rigorously-evaluated one, and that the newest and most sceptical papers in the field have independently begun to document exactly why that is the gap worth filling. Throughout, claims are connected to the evidence already built on this site, because a literature review that also tests the literature is the discipline the whole proposal argues for.

1. Market efficiency and the boundaries of predictability

The intellectual starting point is the efficient market hypothesis (Fama, 1970): if prices reflect available information, systematic return prediction is close to impossible, and the decades-long “factor zoo” of claimed anomalies has largely dissolved under scrutiny. Harvey, Liu & Zhu (2016) show that once the hundreds of published factors are corrected for the sheer number of hypotheses tested, most fail to clear a credible significance bar — a multiple-testing crisis this proposal takes as a methodological warning, not a footnote. Machine learning has sharpened rather than overturned the verdict: Gu, Kelly & Xiu (2020) find that flexible models (trees, neural networks) do extract modest, mostly nonlinear predictability from a large predictor set, but the gains are concentrated and fragile. The proposal engages this evidence rather than wishing it away — its own capstone independently reproduces it, with every model from a linear regression to a transformer tying at a coin-flip on next-day direction. The theoretical home for the proposal’s response is Lo’s (2004) Adaptive Markets Hypothesis, which reframes efficiency as dynamic and regime-dependent: opportunities appear and are arbitraged away at rates that vary with the environment. That reframing is precisely why the proposal targets robustness and decision quality under changing conditions rather than a static predictive edge — a positioning the efficiency literature supports, and one that separates this work from the return-chasing it is often mistaken for.

2. Regime detection and non-stationarity

If markets are adaptive, the central modelling object is the regime. The foundational apparatus is Hamilton’s (1989) Markov regime-switching model, which lets a series’ parameters switch between hidden states following a Markov chain, and its asset-pricing descendants: Ang & Bekaert (2002) show that regime shifts materially change optimal international allocation, and Guidolin & Timmermann (2007) extend this to multivariate regime-switching portfolios. A parallel, more operational line treats regimes as risk states — Kritzman & Li (2010) define a “financial turbulence” index that spikes precisely in the periods risk management most cares about — and the online-detection literature, exemplified by Adams & MacKay’s (2007) Bayesian online change-point detection, asks not “which regime was this?” but “has the regime just changed?” Critically, most of this literature is either offline (smoothing over a completed history, which contains look-ahead) or descriptive (labelling regimes without conditioning decisions on them). The proposal’s contribution is to make regime awareness real-time and decision-facing — filtered, not smoothed; feeding sizing, risk and forecast-calibration — and, unusually, to test whether it pays: its Quant Lab finds that a discrete regime adds nothing beyond a volatility signal for position sizing, yet sharpens risk probabilities and concentrates its benefit at transitions — a nuanced result the descriptive regime literature never had to confront.

3. Volatility: the forecastable dimension

Where returns resist prediction, volatility yields. From Engle’s (1982) ARCH and Bollerslev’s (1986) GARCH to the realized-volatility programme — Andersen & Bollerslev’s (1998) rebuttal of the “GARCH doesn’t forecast” sceptics, and Corsi’s (2009) HAR model — the literature establishes that conditional variance is persistent and genuinely forecastable, if only over short horizons. The natural question is whether that forecastability is tradable, and here the literature is instructively divided: Moreira & Muir (2017) argue that volatility-managed portfolios (scaling exposure by inverse variance) raise Sharpe ratios and produce alpha, while Cederburg, O’Doherty, Wang & Yang (2020) show much of that advantage evaporates out of sample and net of costs. The proposal sides with neither on faith: it reproduces the debate directly, finding that volatility timing buys risk control, not statistically significant alpha, and measures the forecast horizon at roughly a one-month shock half-life. This is the empirical backbone of the whole proposal: the forecastable dimension is risk, so risk — not return — is where the system’s edge is sought. Volatility is priced as well as forecastable: Ang, Hodrick, Xing & Zhang (2006) document that both sensitivity to aggregate-volatility shocks and high idiosyncratic volatility earn abnormally low returns — the idiosyncratic-volatility puzzle — yet the result is notoriously design-fragile (weighting, breakpoints and the related lottery-demand MAX effect all erode it), a cautionary case that underwrites this proposal’s insistence on pre-committed robustness grids over single specifications.

4. Reinforcement learning for sequential decisions

Trading is a sequential decision problem, which makes reinforcement learning its natural formalism. The lineage runs from Moody & Saffell’s (2001) direct (policy-based) reinforcement for trading, through Deng et al.’s (2017) deep direct RL that learns market features and trading decisions jointly, to Jiang, Xu & Liang’s (2017) deep RL framework for portfolio management; alongside sits the execution literature, where Almgren & Chriss (2001) give the canonical optimal-liquidation trade-off between market impact and timing risk. Two weaknesses recur across the trading-RL literature. First, most agents maximise return without hard risk constraints, so a reward-hungry policy will take ruinous leverage for marginal gain — the very failure the risk-sensitive and constrained-RL literature (CVaR objectives, Lagrangian methods) exists to address, and which the proposal’s risk-constrained RL component adopts. Second, most assume stationarity, learning a policy for a market that then changes underneath it. The proposal’s answer is to pair constrained RL with regime-conditioning and a deterministic risk layer that can override the learner outright — and its Quant Lab H6 test confirms empirically what the constrained-RL theory argues: hard, deterministic limits are what turn a fallible learner into a viable strategy.

5. NLP and large language models in finance

Finance has read text for signal since Tetlock’s (2007) demonstration that media pessimism predicts (transient) downward pressure on prices, and Loughran & McDonald’s (2011) finding that a domain-specific sentiment lexicon substantially outperforms general-purpose ones — a reminder that financial language is its own dialect. The transformer era brought domain-adapted models: FinBERT (Araci, 2019), BloombergGPT (Wu et al., 2023) and FinGPT, each improving downstream text tasks. Yet the classic finding survives the technology: the return content of news sentiment is weak and quickly arbitraged, consistent with the efficiency literature above. The proposal draws the honest conclusion — text is not a return oracle — and reframes the News Agent’s job around what text can robustly provide: bounded, auditable, timestamp-safe features (event class, novelty, narrative, risk context), not a buy signal, exactly the stance developed in its schema-constrained extraction and sentiment-and-narrative work. Where most text-in-finance research still chases sentiment-to-return, the proposal treats text as a risk-and-context input under hard controls.

6. LLM agents and multi-agent trading systems — the frontier and its critics

This is where the proposal must be positioned most carefully, because the frontier moved fast. Since 2024 a wave of multi-agent LLM trading systems has appeared — most prominently TradingAgents (Xiao, Sun, Luo & Wang, 2024), which simulates a trading firm with fundamental, sentiment and technical analyst agents, researchers, traders and risk managers, and FinRobot (Yang et al., 2024), an open-source agentic platform — surveyed in a growing body of work (e.g. the 2024 Large Language Model Agent in Financial Trading survey). Their shared design is a swarm of LLM agents deliberating in natural language toward a trade, and their reported backtests are impressive. The proposal is emphatically not a fourteenth entry in this line, and the reason is that the field’s own most rigorous recent work has turned sceptical. Li, Kim, Cucuringu & Ma (2025), with their FINSABER framework, subject LLM investing strategies to broad-universe, two-decade, bias-controlled backtesting and find the reported advantages largely evaporate — and, most tellingly for this proposal, that LLM strategies are “overly conservative in bull markets… and overly aggressive in bear markets,” i.e. they fail to adapt across regimes. Complementary evaluations (2026) stress that these systems’ apparent gains are fragile to evaluation choices and that coordination, not raw model power, governs their behaviour. The literature has thus independently named three gaps: naive multi-agent LLM systems overstate performance, are evaluated unreliably, and are regime-blind. The proposal’s contribution is the deliberate inverse of the prevailing design: agents that do not deliberate in free text but exchange typed features through a structured state (multi-agent coordination); an LLM that is never an execution authority (why an LLM must never trade); regime-awareness by construction rather than as an emergent hope; and evaluation held to deflated-Sharpe, out-of-sample, honest-null standards. Where the field built for autonomy and impressive demos, this proposal builds for control, auditability and evidence — the axis its own critics now say is missing.

7. Backtest overfitting and evaluation methodology

The thread that ties the critique of §6 to the discipline of the whole site is the backtest-overfitting literature. López de Prado (2018) catalogues how standard backtesting silently manufactures false discoveries; Bailey & López de Prado (2014) formalise the correction as the Deflated Sharpe Ratio, adjusting an observed Sharpe for the number of trials, sample length and non-normality; Bailey et al. (2017) quantify the Probability of Backtest Overfitting; and Sullivan, Timmermann & White (2001) show, on the classic case of calendar effects, that data-mined patterns vanish once the search is accounted for. Together with Harvey, Liu & Zhu (2016) from §1, this literature is the reason the proposal treats methodology as a first-class research object. Its Quant Lab is built on these tools — deflated significance, purged/combinatorial cross-validation, point-in-time discipline, and a calendar-anomaly control experiment that reproduces Sullivan et al.’s null on modern data — and reports its own negative results as readily as its positive ones. This is precisely the standard against which the LLM-trading results the critics of §6 dismantled were not held; it is the proposal’s strongest point of differentiation.

Synthesis: the gap this proposal fills

Each of the seven literatures is deep, and none is the contribution. Efficiency and its ML re-examination say return prediction is a near-dead end; the regime and volatility literatures say risk is the structured, forecastable dimension; reinforcement learning offers the decision formalism but usually without hard risk control or regime awareness; the finance-NLP tradition says text carries weak return content but useful context; the new multi-agent-LLM literature offers a compelling decomposition and, through its own critics, a warning that the naive version overstates, mis-evaluates, and cannot handle regimes; and the backtest-overfitting literature supplies the evaluative conscience the rest of the field so often lacks. No existing work integrates these into a single, control-first, regime-aware, risk-governed and honestly-evaluated architecture — one where an LLM informs but never acts, where hard deterministic limits bound a constrained RL learner, where regime awareness is real-time and its value is tested rather than assumed, and where every claim is deflated and reported with its nulls. That integration, positioned modestly as a bet on robustness and decision quality rather than on beating the efficient market, is the proposal’s contribution — and, unusually for a proposal, the majority of its components are already built and tested on this site. This review is the seed of the formal Chapter 2; the notes below develop individual papers in depth.

References

• Adams, R. P., & MacKay, D. J. C. (2007). Bayesian Online Changepoint Detection. arXiv:0710.3742.
• Almgren, R., & Chriss, N. (2001). Optimal Execution of Portfolio Transactions. Journal of Risk, 3(2).
• Andersen, T. G., & Bollerslev, T. (1998). Answering the Skeptics: Yes, Standard Volatility Models Do Provide Accurate Forecasts. International Economic Review, 39(4).
• Ang, A., & Bekaert, G. (2002). International Asset Allocation with Regime Shifts. Review of Financial Studies, 15(4).
• Ang, A., Hodrick, R. J., Xing, Y., & Zhang, X. (2006). The Cross-Section of Volatility and Expected Returns. Journal of Finance, 61(1).
• Araci, D. (2019). FinBERT: Financial Sentiment Analysis with Pre-trained Language Models. arXiv:1908.10063.
• Bailey, D. H., & López de Prado, M. (2014). The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality. Journal of Portfolio Management, 40(5).
• Bailey, D. H., Borwein, J., López de Prado, M., & Zhu, Q. J. (2017). The Probability of Backtest Overfitting. Journal of Computational Finance, 20(4).
• Bollerslev, T. (1986). Generalized Autoregressive Conditional Heteroskedasticity. Journal of Econometrics, 31(3).
• Cederburg, S., O’Doherty, M. S., Wang, F., & Yang, X. (2020). On the Performance of Volatility-Managed Portfolios. Journal of Financial Economics, 138(1).
• Corsi, F. (2009). A Simple Approximate Long-Memory Model of Realized Volatility. Journal of Financial Econometrics, 7(2).
• Deng, Y., Bao, F., Kong, Y., Ren, Z., & Dai, Q. (2017). Deep Direct Reinforcement Learning for Financial Signal Representation and Trading. IEEE Transactions on Neural Networks and Learning Systems, 28(3).
• Engle, R. F. (1982). Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation. Econometrica, 50(4).
• Fama, E. F. (1970). Efficient Capital Markets: A Review of Theory and Empirical Work. Journal of Finance, 25(2).
• Gu, S., Kelly, B., & Xiu, D. (2020). Empirical Asset Pricing via Machine Learning. Review of Financial Studies, 33(5), 2223–2273.
• Guidolin, M., & Timmermann, A. (2007). Asset Allocation under Multivariate Regime Switching. Journal of Economic Dynamics and Control, 31(11).
• Hamilton, J. D. (1989). A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle. Econometrica, 57(2).
• Harvey, C. R., Liu, Y., & Zhu, H. (2016). … and the Cross-Section of Expected Returns. Review of Financial Studies, 29(1).
• Jiang, Z., Xu, D., & Liang, J. (2017). A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem. arXiv:1706.10059.
• Kritzman, M., & Li, Y. (2010). Skulls, Financial Turbulence, and Risk Management. Financial Analysts Journal, 66(5).
• Li, W., Kim, H., Cucuringu, M., & Ma, T. (2025). Can LLM-based Financial Investing Strategies Outperform the Market in Long Run? (FINSABER). arXiv:2505.07078.
• Lo, A. W. (2004). The Adaptive Markets Hypothesis. Journal of Portfolio Management, 30(5).
• López de Prado, M. (2018). Advances in Financial Machine Learning. Wiley.
• Loughran, T., & McDonald, B. (2011). When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks. Journal of Finance, 66(1).
• Moody, J., & Saffell, M. (2001). Learning to Trade via Direct Reinforcement. IEEE Transactions on Neural Networks, 12(4).
• Moreira, A., & Muir, T. (2017). Volatility-Managed Portfolios. Journal of Finance, 72(4).
• Sullivan, R., Timmermann, A., & White, H. (2001). Dangers of Data Mining: The Case of Calendar Effects in Stock Returns. Journal of Econometrics, 105(1).
• Tetlock, P. C. (2007). Giving Content to Investor Sentiment: The Role of Media in the Stock Market. Journal of Finance, 62(3).
• Wu, S., Irsoy, O., Lu, S., Dabravolski, V., Dredze, M., Gehrmann, S., Kambadur, P., Rosenberg, D., & Mann, G. (2023). BloombergGPT: A Large Language Model for Finance. arXiv:2303.17564.
• Xiao, Y., Sun, E., Luo, D., & Wang, W. (2024). TradingAgents: Multi-Agents LLM Financial Trading Framework. arXiv:2412.20138.
• Yang, H., Zhang, B., Wang, N., et al. (2024). FinRobot: An Open-Source AI Agent Platform for Financial Applications using Large Language Models. arXiv:2405.14767.

This review is a working document developed in the open, not the final Chapter 2. Classic finance/econometrics citations (efficiency, regime-switching, GARCH, RL for trading, finance-NLP) are foundational and stable; the 2024–2026 multi-agent-LLM and ML-asset-pricing citations were verified against current sources at the time of writing, and details (volumes, venues) should be re-checked at formal write-up. Where the review asserts an empirical claim about markets, it is tested on this site rather than taken on trust.