Quant research will automate experiment production before research ownership

Quant research is exposed to a sharp fall in the cost of producing an experiment. We expect the standard sequence of data preparation, feature construction, model fitting and report generation to become increasingly agent-run. The scarce contribution moves toward choosing questions worth answering and deciding which results survive contact with a market.

Our dates below are DXRG scenario assumptions, written in October 2026. Jane Street's quantitative researcher description identifies experiment design, financial datasets and model development as parts of the role. It also describes researchers working with engineers and traders. That task bundle gives us a useful reference for decomposing automation; the company supplies no automation timetable.

The first transition: more experiments per researcher

For 2026-2028, our central scenario is supervised automation of well-specified research work: reproduce a published method, build a dataset from approved sources, run a frozen evaluation and explain why a candidate failed. Here, mid-horizon prediction means forecasting outcomes several hours to several days ahead using information available before the forecast. Public news and conventional price features are examples of inputs, with timestamp and licensing constraints preserved.

The mechanism is straightforward. A research agent can write code, execute it, inspect errors and revise the implementation without handing every intermediate step back to a person. The METR long-task paper found improving autonomous performance on software tasks, with a historical horizon doubling trend of roughly seven months. Its conditional extrapolation concerns software work, rather than the economic value of financial discoveries. We use it as evidence that multi-step execution can improve, while treating finance transfer as an assumption to test.

A researcher could ask for a baseline comparison and receive the code, rejected variants and reproducible results together. That changes the unit of work from a notebook edit to a completed experimental package. We have higher confidence in this transition for environments with clean interfaces and inspectable acceptance criteria. Ambiguous labels, missing timestamps and repeated human rescue reduce confidence even when a final report looks polished.

The near-term threat to routine research labor is therefore real. If one researcher can review many complete packages, the human hours needed per conventional experiment can fall substantially. Firms may respond by investigating more markets rather than shrinking teams. Either response changes which skills deserve investment.

The harder transition: selecting discoveries

For 2028-2031, our scenario extends autonomy to a bounded research program. An agent proposes hypotheses within an authorized data universe, budgets compute, tracks failed trials and delivers a portfolio of candidates to independent review. This is a much stronger projection because choosing the experiment is harder to verify than executing it.

Consider an illustrative program testing whether a public announcement predicts next-day returns. An agent can quickly discover a promising transformation. It can also quietly inspect enough variants to overfit the evaluation window. Speed increases both possibilities. The program needs an untouched later-period evaluation, an accurate record of the search and a cost model that includes execution. A better statistical score establishes predictive skill; a feasible after-cost decision establishes a different result.

Our confidence rises if agent-generated programs repeatedly beat an equal-budget human baseline on previously unseen market periods, with the same permitted inputs and candidate-selection rules. It falls if gains vanish when failed searches count, time availability is reconstructed or turnover costs are included. Those observations should move the scenario dates, rather than becoming exceptions hidden in the prose.

For 2031 onward, full ownership of selected research programs becomes plausible under continued capability growth and dependable review infrastructure. A single calendar date for the entire researcher role remains a poor forecast. New instruments, regime changes and internal coordination constantly alter what the job contains.

Better automation can strengthen incumbents

Commodity public-signal work becoming cheaper would open opportunities for small teams. It could also reward firms that already possess good research infrastructure. Jane Street's role description reports extensive data and compute resources. We infer that those assets can support more parallel experiments and stronger internal validation, conditional on effective implementation. Capital, market access and proprietary observations remain separate advantages.

Our strongest projection is that manually producing standard experiments becomes a weaker basis for professional differentiation this decade. The useful test is whether an agent can deliver a genuinely new, reproducible result under frozen evaluation conditions. A lab should measure that transition directly before treating a faster research assistant as an autonomous discoverer.