A backtest is the best evidence a small investment team has before it commits capital, and it is easy to fool. Figures revised after the fact, a universe missing the companies that later failed, or a result carried by a handful of stocks can each make a strategy look better than it is. Once the strategy is live, teams often learn that it has drifted from its backtest the slow way, from the returns.

I build research tools for investment teams: backtests that resist those errors, monitoring that shows when a live strategy leaves the range its backtest predicted, and alerts on the measures your team watches. I’ve done backtesting work for an investment manager, and before software became my main work I ran fundamentals backtests of my own, keying every query on when a figure was published rather than the period it describes. The strategy monitor, the backtest check and the macro alerts screen in my research desk demonstration show the ideas on simulated data.

Backtests that hold up

  • Each figure is dated by when it became available, so the test uses only what was known on each date. I explain how in The timestamps that make financial data usable.
  • The universe includes companies that were later delisted or acquired.
  • Trading costs follow the turnover the strategy actually produces.
  • The same ranking runs on separate slices of the universe. If one slice carries the result, it shows.
  • A stretch of history is held back until the rules are fixed, because a result you can’t measure out of sample isn’t evidence yet.

A backtest is finished when anyone on your team can rerun it from the code and the data and get the same numbers, with every assumption written down.

Watching a live strategy

Once a strategy is live, the question is whether it still behaves like its backtest. A monitor compares live returns and turnover with the range the backtest led you to expect, tracks each slice of the universe separately, and flags the strategy when it leaves that range.

Alerts on what you watch

Many teams check the same few measures every morning, such as market breadth, a recession indicator or a spread. I turn those into scheduled checks with alerts, using your definitions rather than mine. Each alert runs on history first, so you can see how often it would have fired, and whether it fired on figures that were later revised.

When a custom tool is worth it

Research platforms bring a great deal of data together, and for open-ended questions they are the right tool. A tool built for your team makes sense in narrower cases:

  • The work is a process your team repeats. A platform lets an analyst run any analysis once; a custom tool runs your rules on a schedule and keeps a record of every run, so this month’s result can be checked against last month’s.
  • The data includes what no vendor can see, such as your positions and your analysts’ notes.
  • The answer needs several providers at once. Each provider’s tools work best on its own data, and combining them usually falls to a spreadsheet someone updates by hand.
  • You already license the data you need, or it comes from public sources you’re free to use, such as SEC filings. Platform seats are priced per person, every year. For a small office, a few seats can cost more in one year than building a tool on that data and keeping it running.
  • Your rules should stay inside the firm. They live in your own repository and run on your own systems, not saved on someone else’s platform.
  • You want to own it. The code and its documentation are yours, and the settings your team changes most often, such as thresholds and queries, live in files an analyst can edit without changing the code.

When the questions change from week to week, a research platform is usually the better buy.

Working together

Building a monitor or an alert system is a fixed-scope project that ends with a handover: the code, its tests, instructions for running it, and a walkthrough. It is finished when it has caught the historical cases it should and its first live runs match what your team sees by hand. You provide the data you’re licensed to use and someone who knows the strategy’s rules. Your strategies and data stay yours; nothing about them appears in my examples or writing. This is engineering, not investment advice: the strategies and the decisions remain your team’s.

Message me on LinkedIn with the strategy or measure you want to test or watch, and the data you have for it.