Backtests versus the track record

Lab is a safe place to ask "what if". It places no trades, changes no setting and accepts no configuration: the only things you can carry out of it are a saved scenario and a draft guidance note you choose to open on Strategy › Guidance. The title carries one sentence for the tab you are on and, at the top right, which price history every figure came from. On Stress test a strip under the tabs shows your book and the two limits that pause it, so the result is read against the same rules that run live.

Someone whose role on the book is read only sees the words "Lab · read only" in place of the tabs: what-ifs and backtests use the owner's computing budget, so they are for the owner, an operator or a reviewer. Results that were saved stay readable on Performance.

The four tabs

Stress test applies a what-if move to your current positions: the whole market, one name, a sector, or wider swings. It shows today after the move, the room left before the daily halt, how far under the book's best-ever value it would sit, and which positions would be through their stop or their thesis invalidation. Results are instant.

Replay walks today's book day by day through a real episode with the same halt rules. It shows what holding through would have done, what selling at the halt would have done, the difference, and the first day the system would have stopped trading.

Projection resamples real daily moves into hundreds or thousands of paths over 3 to 24 months, with halts applied on every run and AI cost taken off. A compare switch runs two rule sets side by side.

Backtests: queued, persisted, honest

The Backtests tab is a list of past runs, a page at a time, with a New run button; the Sharpe explorer is the section at the end of the same tab. The list's header counts the runs that are listed, and each row says what the run was in a sentence and carries five figures: the yearly return, the return against the S&P 500, the worst fall, the Sharpe and the chance the edge is real. New run opens the builder in place of the list: the form on one side, and on the other what will run, how many runs that is, about how long they take and what they cost. Three kinds of test:

• One run: one engine run of the mix you choose (any registered algorithm with a weight, the stocks, the period, costs, sizing, how often it rebalances and the limits every simulated order must pass).
• A grid of settings: vary one or two settings over a list of values and run one backtest per combination. The grid's own page reports every setting tried, the best one, and the reasons to doubt a winner picked from that many tries.
• A scripted week: replay the PM day by day through the desk's safety limits. The scripted stand-in is free and repeatable; letting Claude manage the account needs a model key set up for test runs and a spend you confirm. It is a behavior test, never evidence of returns, and every row says so.

Runs wait their turn and keep going while you are away. Choose a row and the run opens as a page of its own, with a breadcrumb (Lab, Backtests, the run's name) that takes you back to the list; while it runs, the page checks every few seconds and shows the result the moment it finishes. Every result carries Reasons to doubt this result, the figures that argue against the headline: Could it be luck? (the deflated Sharpe: the chance the true Sharpe beats what the luckiest of N skill-less configurations would show; below about 95% the result cannot be told from luck in choosing), First part against last part (the first 70% of sessions against the last 30%; nothing was fitted on one and tested on the other, it shows whether the result leans on one era), What a single year could look like (the returns reshuffled into thousands of possible years, with how many lose money and how many beat the S&P 500), and for grids Did the search pick noise? (the probability of overfitting) and On years it had not seen (a walk-forward line where each year's setting was chosen on the prior years only). Tick Compare on two finished runs and choose Compare the two to read them side by side.

Nothing you find here changes the blend. To act on it, go to Signals and move the weights yourself.

The track record

Performance is the other side of the ledger: what the book actually did in paper, after AI cost if you flip the switch. It has five tabs. Results chains daily returns (deposits stripped out) into an equity line against the same money in SPY and a 60/40 mix, draws the chart of drawdowns and grids returns by month. Trades keeps the Trade journal of every round trip with win rate, expectancy per trade and profit factor. Track record lists every proposal the desk made and what happened next, Statement is one calendar month at a time, and Evidence holds the proof so far and the backtest report. The equity line starts on the first funded day; the journal, win rate and breakdowns start with the first position bought and fully sold. There is nothing to flatter before then.

The first days of a book

A new book has a handful of trading days, and the chart says so: while it has fewer than ten, every close is drawn as a point (the book and both benchmarks), the points are joined by straight segments, and the chart names the count in the same words as the summary above it ("4 trading days since the Sep 11 start"). The segments are straight on purpose. Each point is one day's close and nothing was measured between two closes, so a smooth curve through them would invent prices that never printed. The first point is the close the returns are measured from, and the summary line names it ("measured from the Sep 11 close"). During the trading day the newest point is today so far, drawn as an open ring, and the chart says it is not a close yet. From the tenth trading day the points give way to the plain line.

Some figures are withheld until there is enough to measure. Sharpe, Sortino and volatility need twenty trading days; beta and correlation against SPY need twenty trading days both series cover; the win rate, the average result per trade and the profit factor need five closed trades. Until then the tile says what it is waiting for and how far along it is ("shown after 20 trading days", "4 so far"; "3 of 3 won so far") and never prints a number, because a ratio off four days or a win rate off three trades is noise dressed as a measurement. On a one-week view of an older book the same tiles read "needs 20 trading days in the range": the week is short, not the book. The return, the profit and loss, the worst fall from a high and the AI cost are facts from day one and always print.

The benchmarks are compared over the same days as the book. SPY's and the bond fund's official closes are published a few hours after the bell, so between the close and that moment the page uses that day's last prices for the newest day, draws that point as an open circle and says so under the chart with the time. If a benchmark close is missing altogether, your newest close is still drawn (dotted, and the line under the chart says it is final and only dotted for want of a comparison), the comparison stops at the last day both have, and "vs SPY" states the range it covers. A benchmark the platform only holds from a later day starts there, from the book's own level that day, and the line under the chart names the day.

Research adds a second live measure: every brief's checkable calls are scored from daily closes at the horizon, with no model grading, and Research shows hits over scored calls per brief (no combined hit rate across briefs yet).

A backtest tells you what a rule would have done on the winners of the last 20 years. The track record tells you what this system did, with this data, this month. When the two disagree, believe the second. See what counts as proof.