The tidal wave of AI research in finance

The question in financial AI is changing quickly. It is no longer enough to ask whether a language model can read a filing or produce a forecast. Researchers now need to ask which representation is doing the work, how expensive it is to deploy, how long its signal survives, and whether the result can be reproduced after the market and the model have moved on.

That is why the current wave of research feels less like a single new technique and more like a change in the research object itself.

A result worth taking seriously

In a recent paper in the Journal of Financial Economics, authors show LLM-based sentiment strategies outperform traditional factor-model strategies on Sharpe ratios. The result is not simply a victory for the largest decoder-only model: older and much smaller embedding models also perform surprisingly well.

That second result matters. It suggests that the useful object may not be a model that writes fluent prose. A compact model that maps a document into a stable representation can be enough to recover economically relevant structure. The representation can then feed a portfolio rule, a factor construction step, or a risk model without paying the full computational cost of generation.

The finding should still be read as an empirical claim with boundaries. Sharpe ratios are sensitive to the sample, portfolio formation, turnover, transaction costs, leverage, risk-free-rate treatment, and the distinction between an in-sample fit and a genuinely point-in-time forecast. A high ratio is a reason to inspect the information set and the backtest design, not a substitute for either one.

The model can stay fixed while the opportunity decays

The most interesting pattern is temporal. Holding the model constant, performance declines as the evaluation period moves forward. One interpretation is market efficiency: entrants learn the same signal, competition arbitrages away the opportunity, or newer models make the old representation less distinctive. In that reading, model age is not just a technical detail. It is part of the economics of the signal.

There are other possibilities. The market regime may have changed, the asset universe may have drifted, the news distribution may have changed, or implementation costs may have risen. A constant model does not by itself identify the mechanism. It does, however, give researchers a clean research question: how much of performance decay comes from the market, how much from the corpus, and how much from the representation?

Beyond the Copilot

For many businesses, the first focus of AI adoption has been to aid, augment, and automate existing processes and knowledge work. Companies with long-standing customer relationships, proprietary tools, and institutional knowledge need to retrieve, organize, and enrich that knowledge before a model can use it as context. That retrieval layer has become an important - and difficult - starting point for generative AI adoption.

For these businesses, access control, hallucination, cost, expertise, scale, and maintainability remain important obstacles. Models provide a simple interface for prototyping new workflows, but long-term memory, factuality, and brand consistency remain non-negotiable for real-world value. In practice, two model families are especially relevant: autoregressive decoder-only models used for generation, such as ChatGPT, Claude Code, and OpenClaw; and encoder-only models used to produce representations for search, clustering, and the retrieval stage of retrieval-augmented generation (RAG).

Decoder-only models are powerful because they can synthesize, explain, and act across open-ended tasks. Their autoregressive generation also creates challenges for reproducibility, latency, cost, and auditability. Production systems therefore often need retrieval, structured tools, deterministic checks, and other controls around the model to make its outputs verifiable.

Encoder-only models are narrower, but they are usually cheaper and faster at producing stable vectors. Those vectors make relationships among documents inspectable through similarity, clustering, and nearest-neighbour structure. The representation is not automatically interpretable, but its behaviour is easier to measure repeatedly: a researcher can apply the same encoder to a dated collection of filings, compare firms in a common space, and preserve the resulting vectors as an auditable research input. For finance applications, this distinction matters because the model is often being used as a measurement instrument rather than as a conversational interface.

In The Embedder’s Dilemma, Adnan El Assadi, Niklas Muennighoff, and Jinhyuk Lee compare ten LLMs with 26 dedicated embedding models across 37 tasks. Aggregate quality is nearly tied, but the strengths differ: LLMs lead on reasoning-heavy retrieval, embedding models lead on classification, and the two paradigms are roughly matched on clustering and semantic similarity. The cost gap is substantial: for the benchmark studied, an LLM can cost up to 1,431 times more than an embedding model of comparable quality, while the measured speed penalty varies by model. The lesson is not that embeddings always win. It is that quality, latency, cost, and task fit have to be evaluated together.

That comparison gives financial researchers a practical design rule. Use a generative model when the task requires synthesis, explanation, or multi-step reasoning. Use an encoder when the task requires the same transformation to be applied repeatedly across a large, dated document panel. The latter is common in finance: filings, earnings calls, news, and research notes become firm-level characteristics that feed portfolio sorts, forecasts, factor models, or risk measures. In those settings, model age, inference cost, and representation stability are part of the empirical design - not implementation details to be hidden behind a chatbot interface.

The journal wave

The publication market is signalling the same change. The Journal of Banking & Finance is inviting work on “New Technology and Economics for Banking and Risk Management”; the Journal of Financial Economics has a call on the “Economic and Financial Implications of Artificial Intelligence”; the Global Finance Journal is soliciting papers on “Artificial Intelligence and Capital Markets”; and the European Journal of Finance is seeking work on “Generative AI and LLM in Financial Modelling and Applications.”

That concentration of calls is a useful signal about where the field is going. It also raises the standard. The next generation of papers will need more than a model name beside a Sharpe ratio. They will need a dated information set, meaningful baselines, and a clear account of what the representation contributes.