Quantitative research internships: build evidence, not an impressive backtest

Why 90% accuracy might show no improvement

Original fictional example: an untouched test set has 100 days. The market rises on 90 and falls on 10. Your model predicts 'up' every day and reports 90% accuracy.

A baseline that always predicts 'up' also achieves 90%. On this measure, your model has added nothing. The result does not establish profitable trading either: accuracy says nothing about the size of the gains and losses, position sizing or costs.

Before fitting another model, write what would count as improvement and why that measure matches the question. If you are predicting the direction of a return, classification performance is one check. If you are proposing a trading rule, its payoff and risk need separate evaluation. Do not switch metrics after seeing which makes your result look best.

Protect the test from information you would not have had

For a time-dependent project, record the prediction time and the timestamp at which each input becomes available. Tomorrow's closing price cannot be an input for a prediction made today. A value published later but labeled with an earlier reporting period can create the same problem.

Set aside a final period before choosing models. Use earlier data to develop and compare alternatives. If you repeatedly inspect the final period and revise the model, it has become part of development; do not continue describing it as untouched evidence.

A second fictional check concerns costs. Suppose a strategy averages 20 basis points gross per trade, where one basis point is 0.01 percentage point. Subtracting 18 basis points of modeled costs leaves only 2. An additional 6 basis points of execution cost changes the result to −4 basis points. All figures use the same per-trade return basis. The gross chart alone would hide that sensitivity.

• Keep the simple baseline and the same evaluation dates beside each model result.
• Log unsuccessful experiments, not just the best run.
• Document data revisions, missing records and costs so another reader can reproduce the comparison.

Present the failed idea as carefully as the successful one

A useful notebook has one defined question, a baseline, a defensible split and an explanation of the errors. A short accompanying note can say what survived testing, what disappeared and what remains unknown. Use licensed public data or clearly labeled synthetic data rather than private employer records.

Practice answering: 'What information would have been available at the time?', 'How many alternatives did you try?' and 'What would make you reject this result?' Those questions reveal the quality of the evaluation more directly than listing algorithms.

Ask the employer whether the intern's expected output is a research report, a model or an improvement to the research system. Tailor the sample to that work. A careful nonfinancial statistical project can demonstrate these methods without being relabeled as a trading strategy.

Before you apply

Eligibility

Compare the degree requirements on each position. Undergraduate, master's and doctoral research routes are not interchangeable.

Timing

Citadel Securities describes an 11-week internship with a separate research track. Current role pages determine application availability.

Location

Read the research team's office requirement; access to online datasets does not establish a remote employment arrangement.

Pay

Compare offers at the same degree level and over the actual paid term, not an experienced researcher's annual compensation.

Sources: Citadel Securities internships · Jane Street internships

Job titles to search

Quantitative research intern

Form a hypothesis, test it and explain whether the evidence supports using it.

Systematic research intern

Study a repeatable investment or market process, including how a proposed rule would fail.

Research engineering intern

Make data processing and experiments reliable; the output may be infrastructure rather than a new model.

Where to look

Search routes, not confirmed vacancies.

Compare the work and try an example

Use a role comparison or worked exercise to choose your preparation before tailoring an application.

• Corporate treasury versus FP&A internships: cash, forecasts and decisions
• Financial due diligence internships: what to analyze and how to practice
• Private credit versus corporate banking internships