Our approach
Every idea that reaches a portfolio has been written down in advance, tested on data it could not see, reviewed by a colleague and sized by risk.
Overview
An idea has to survive four people before it reaches a portfolio: the researcher who proposes it, the colleague who tries to break it, the risk manager who sizes it and the committee that approves it.
Six stages
The same stages apply to a credit view, a client allocation and a trading signal. Only the evidence differs.
The researcher writes a one-page memo: the expected effect, why it should exist, the horizon, and the result that would prove it wrong.
Data engineering confirms that every input is point-in-time, licensed for the use, and documented from source to field.
The idea is tested on a development sample, with costs included. A holdout period stays untouched until the final check.
A colleague who did not build the work reproduces it from recorded code and data, and tries to find reasons it will fail.
The Investment Committee approves, rejects or sends the idea back. Risk sets its initial size and limits before any capital is used.
Live results are compared with the original memo each quarter. Ideas that stop behaving as described are reduced or retired.
Independent review
Reviewers work from a written checklist and sign their name to the result. A failed check sends the work back, not forward.
Results match when rerun from the recorded code version, data snapshot and parameters.
The holdout period was not used to choose any parameter, and the log shows when it was first opened.
Every variation tried is counted, so a lucky result among many attempts is not mistaken for a real one.
Transaction costs, financing and capacity limits are modeled at the size the portfolio would actually trade.
Data
Datasets are stored as they were known on each date, including later revisions, so a backtest never uses information that was not yet available.
Each field records its vendor, load time and transformations, so any number in a research result can be traced to its source.
New datasets are reviewed by compliance and legal before use, covering licensing, how the data was collected and whether it could contain material non-public information.
Evidence by discipline
| Discipline | Typical idea | Evidence required | Signed off by |
|---|---|---|---|
| Fixed Income | Overweight a credit sector for the next twelve months | Issuer-level fundamentals, spread versus fair-value model, downside under a recession scenario | Head of Fixed Income; Chief Risk Officer on size |
| Asset Management | Change a multi-asset portfolio’s strategic allocation | Effect on the client’s stated objective, drawdown and liquidity; tax impact for separately managed accounts | Head of Asset Management; Investment Committee |
| Quantitative Trading | Add a new systematic signal | Holdout results after costs, number of trials, capacity estimate, correlation with existing signals | Investment Committee; independent model validation |
Published research
Why a holdout period only works if nobody looks at it, and how we keep it that way.
Read the paperHow revised and restated data can make a backtest look better than reality.
Read the paperWhere machine learning helps our researchers, and the checks it must pass first.
Read the articleDescriptions of our research process are not a guarantee of results. Quantitative models may not perform as expected.