AI for quantitative research
Backtest signals with the bias checks a validator would run, and review exposures and limits before the investment committee. Attribution and stress reports come from the returns, holdings and risk model exports already on your desk.
momentum-12-1-memo.docxExampleResearch Memo: 12-1 Price Momentum in US Large Caps
Result
A dollar-neutral portfolio, long the top decile and short the bottom decile of 12-1 momentum and rebalanced monthly, returned 6.1% a year gross and 4.0% net, with 16.8% volatility and a net Sharpe ratio of 0.24. Costs of 5 bp per side on monthly turnover of about 140% of each leg, plus 40 bp a year of borrow on the shorts, take 2.1 points a year. The worst drawdown was −47%, in the 2009 momentum crash.
Bias checks
The universe is the top 500 US stocks by float-adjusted market cap, rebuilt each month from point-in-time data, and delisted names carry their delisting returns, with −30% assumed where none is reported. Signals use returns from month t−12 to t−2, so formation never sees the holding month. Haircutting for the three formation windows tested, 6-1, 9-1 and 12-1, lowers the net Sharpe from 0.24 to 0.19.
CheckedOpen: test a volatility-scaled version before this goes to the PM, given the 2009 drawdown.
How it works
- 1
Put the research in one folder
Price history, point-in-time constituents, fundamentals, holdings snapshots, risk model exports, benchmark files, prior memos and the backtest scripts your team already trusts.
- 2
Ask for the study or the pack
It writes and runs the code and hands back the memo, workbook or report. It also states the universe dates, signal lag and cost model it used, and names the bias checks it ran or skipped.
- 3
Break it, then decide
Rerun the timing on a few dates by hand and press on the trial count and the cost assumptions before you read the counterevidence section. Every number cites its source file, and the signal stays yours to accept or kill.
Ask it for
“Backtest a 12-1 price momentum factor on point-in-time S&P 500 constituents, charging 5 bps half-spread and 40 bps annual borrow on shorts.”
momentum-12-1-memo.docx“Compute IC, ICIR and quintile spreads for the gross margin change factor, first on the training window and then on the held-out 2022 through 2024 sample.”
factor-ic-quintiles.xlsx“Review August holdings against the mandate and flag positions contributing over 4% of TE variance, active factor bets over 0.5 sd, and issuers above 8%.”
risk-review-2026-08.xlsx“Run monthly Brinson attribution for the US large-cap composite across Q1, Carino-link the months, and reconcile the effects to the 2.1% excess return.”
Q1-brinson-attribution.xlsx“Apply the Fed’s 2026 severely adverse scenario to the general account portfolio and write up the results with the equity, spread and unemployment paths.”
GA-severely-adverse-2026.xlsx“Mark up the analyst’s momentum backtest for look-ahead bias, survivorship, missing delisting returns, cost assumptions and trials, and list every failed check.”
backtest-review.pdf
Inside the Quantitative Research mode
Quant research memo
Drafts the signal memo with the hypothesis, universe, timing rules, cost model, bias checks and the counterevidence that would kill the result.
Portfolio construction and risk review
Builds a pre-committee review from a holdings file: contribution to tracking error by position, factor exposures, ex-ante tracking error and limit breaches.
Performance attribution and GIPS
Assembles the monthly attribution pack, Carino-links the periods, and documents the composite work so your verifier can reperform it.
Stress testing and scenario analysis
Runs historical and hypothetical scenarios on the portfolio and writes the stressed results with the assumptions behind each number.
- Office skills
- Excel, Word, PowerPoint, PDF
- Tools
- Read, edit and write files, Search files, Run code, Search and read the web, View images
- Connectors
- SharePoint, Google Drive, Gmail, and additional custom connectors
- House style
- Your fonts, layouts and rules, on every file
Who it’s for
Quantitative researchers, quant strategists and systematic portfolio teams at hedge funds, asset managers, insurers' investment departments and bank strategy desks.
Questions
Will it introduce look-ahead bias into my backtest?
It can, which is why this mode treats look-ahead as the failure you plan for. It writes the timing rules into the code, so a signal uses data through t-1 and the position is opened on the next bar. Universes come from point-in-time constituent files where you supply them, and delisting returns are included when the data supports them. The memo lists the bias checks it ran and the ones it could not run. You review the code and results before any number leaves your folder.
Can I trust a Sharpe ratio the model hands back?
Not on its own, and you should not fully trust one from a notebook when the trials behind it are hidden either. Costs in bps, turnover, results by subperiod and how the number moved across regimes are all in the memo, so a reader can discount the result. Every trial variant you told it about is included, even the ones that failed. The memo renders no verdict, because the judgment on whether the signal is real is yours.
Can we claim GIPS compliance based on its composite reports?
No, GIPS compliance is a claim your firm makes firm-wide and maintains, not something a file confers. Give it your composite definitions and the GIPS rules, and it builds composite returns, dispersion and attribution reports against them. Each figure cites the input files behind it, and data it could not reconcile is flagged. The compliant presentation is yours to assemble, and a verifier examines only whether the policies behind it were followed.
Will our model validation team accept AI-assisted research?
Validation is independent by design, so the validator decides, not the developer and not the tool. The mode gives them a package they can reperform: the hypothesis, the data files with their dates, every modeling assumption, results with sources, and a record of the checks that failed. Under model risk management guidance like the Fed’s SR 11-7, credibility comes from documentation a second person can reproduce.
Where does our data go when the model processes it?
Ultramarine Work changes files only in the project’s folder. When a request needs it, it can read files elsewhere on your computer, never passwords, keys or app data, and you can turn that off in settings. The model runs on Paper Instruments’ servers, so the content it works on is sent there to be processed. Paper Instruments is SOC 2 Type II, and our security and compliance details are on our compliance page.