Running studies

A backtest says what a rule would have done. A study says what you claimed before you looked, and what the history did to that claim. This chapter is about the second kind, and about the three checks every study on this site goes through.

Validation

The three checks

Time Travel
Pin a virtual today. Only what existed on that date is visible: the constituent lists as they stood, the filings as first published, the prices up to that day. The strategy is built with what was knowable then and replayed forward bar by bar, so a design cannot borrow from the future by accident.
Walk-forward
Fit on one span, test on the next, move both forward, repeat. The out-of-sample pieces stitched together are the result you report; the fit on the span it was tuned to says nothing about the next one.
Autopsy
Where the result came from, period by period. A gain that lives in one year is a different claim from one spread across twenty, and Autopsy is where you find out which you have before anyone else does.

All three are on the Practitioner plan and above. A backtest on the free plan is still a real run on real history; it is the checks that a plan adds.

From hypothesis to paper

The study cycle

  1. Build the circuit on the canvas and run the consistency check. The engine's prospectus states what the circuit will buy, when, and for how long. That reading is what the registration freezes.
  2. Register the hypothesis at the anchor date: what you expect, the arms if two rules compete, and the family, every variant you intend to try, so the paper can count them.
  3. Run the window, log the run, move the anchor to the next window. A window's report is frozen when it is logged; a later change to the circuit starts a new study.
  4. Compile. The paper is written from the frozen reports: the tables, the figures, and every window's report behind them.
  5. Submit. The platform reviews it. A published paper goes into the research library, free for anyone to read, with the circuit that produced it.

Practitioner registers a hypothesis and keeps the record in QL-Git. Research walks, compiles and publishes. QuanterLab Desktop runs the same cycle on your own machine, with your own compute, and uploads the compiled study to the platform as a draft for the same review.

Counting the tries

Multiple testing

Ten variants tried and one reported is the commonest way a backtest misleads its own author. Here the family is declared at registration, the paper shows every member on one figure, the arms that died included, and the headline statistic is deflated for the number of variants that were tried. A reader of the paper sees the count, and so do you.

The bar

What a paper has to carry

The declared hypothesis and the family size on the page. What is bought, when and for how long, before any number. Costs charged as real where the study models them, and the ladder it used named. The limits stated once. Enough alongside the paper that someone else can re-run it. The paper skill is the same list, written for an assistant that drafts one.

Documentation

The chapters

An account, the modules, a first backtest, a first circuit, and where your work is kept.
Time Travel, walk-forward and Autopsy; the cycle from a registered hypothesis to a published paper.
Quantin, Claude, ChatGPT and the desktop's seat: what each one can do, where it runs, and what account it needs.
The connector URL, the setup steps for claude.ai and ChatGPT, the four skills and the tool list.
Where the market data comes from, what leaves the platform, and the daily allowances.
The three plans and what each adds, moving between them, firms and institutions, and the legal pages.