Sweeps

A sweep runs your strategy once for every combination of the settings you choose to vary, over the same months, and ranks them. It is the quickest way to learn which settings matter and where the strategy's good results cluster. Its weakness is built in: the winner is picked and graded on the same months, so a sweep's best result is always flattered by the search, and the results say by how much.

#Before you start

• A project with a base run: a finished backtest with trades, on Build.
• Enough trades to compare combinations: a search needs at least a few dozen, and many more to compare many combinations.
• Credits: a sweep costs a backtest's price for each combination it runs.

#Plan and run a sweep

1. 1

   Open the plan

   On the project's Optimize stage, press Plan a sweep. Once the project has searched before, use the Plan button beside the method switch, with Sweep chosen.

   You should see the Sweep plan opens in place of the canvas, drafted by the copilot from your run.
2. 2

   Choose what to vary

   Under What to vary, keep, edit or remove the copilot's rows, or add settings with Test it. Every row should include today's value, so the sweep can tell you whether anything beats what you run now.

3. 3

   Check the yardstick

   Under How the winner is picked, Ranked by starts at your goal's default. If this is the project's first search, the choice locks when you launch.

4. 4

   Keep the hold-back

   Under the window, holding back the last months is on. Leave it on if you might walk the result forward, so the walk-forward is graded on months this sweep never saw.

5. 5

   Check the budget and the price

   The bar shows how far the plan stretches its trades, from Fair to Mostly luck, and its price. Trim a plan that reads Mostly luck.

6. 6

   Run it

   Press Run sweep, name it, and confirm the price.

   You should see its row appears under Searches, filling in as combinations finish.

#Read it

The sweep's results open under its row, on the Manual Bulk tab: every combination in a table, and an analysis of the whole sweep with its verdict, whether its winner looks like real edge or selection luck. Reading Sweep Results goes through it.

Then decide:

• It found something real. Promote the best combination to be the project's champion, or first test its ranges with a walk-forward. See Champions & Promotion.
• It found a plateau. Open the plan again: it starts from what the sweep said, and offers to zoom in on the plateau or walk the ranges forward, each priced.
• It found noise. Change the strategy in Build. More searching on the same months only finds luckier noise.

#If it doesn't work

• The plan won't run: too few trades. A search needs at least 30 trades to learn from. Run a longer backtest first.
• Some combinations never ran. Their row says why, and their share of the credits is refunded.
• Every combination stopped early. The engine's safety limits stopped each one, so there is nothing to rank. Fix the strategy on Build and run it again.
• There's no Plan a sweep. The project has no base run yet: run a backtest on Build first.