Optimiser

A study runs the strategy over many combinations of its settings and reports how stable the best one is. The value here is the shape of the results, not the single best row.

Before you start

Check the settings you want to vary, give each one a step, and the form says how many backtests at most that allows, how many run at once, and, once a study has finished at that pace, how long the next one should take. The budget fields carry plain names: starting samples, search rounds, settings per round, neighbour check, run at once.

Choose the build too. An empty picker means the newest build at each evaluation's run time. A build that is no longer on disk is refused by name before the study starts, and every evaluation records the build it ran.

Smart search and full grid

Smart search walks the space in rounds and spends its budget where the result looks promising. Full grid runs every combination of the step values of the checked settings exactly once, with no random start and no rounds, which is the honest choice when the grid is small enough to enumerate. Every checked setting needs a step for a full grid, the form counts the runs before you commit, and more than two thousand combinations is refused.

The verdict

The report leads with one plain sentence about the shape of the results: a robust hill, a soft hill, a spike that is likely luck, or no positive region at all. Then the pick, named in the strategy's own parameter names, a stability score from zero to one hundred, whether the pick is significant, and how likely the result is to be luck from trying many settings.

Under the pick, one line says how it was chosen. The pick sits on a smoothed model of all the results, subtracted by how much the area varies and how unsure the model is there, rather than being the highest single row. The table still names its own best row, and marks in green when the two agree and in gold when they disagree, so a spike can never be sold as a plateau.

A surface view draws the whole area with a colour scale titled by the metric, the pick as a white diamond and the highest raw row as an orange cross, and you can switch the pair of settings on the axes.

Taking a setting forward

Click a row to open that setting's own backtest. A row's sweep action, and the verdict's own sweep button, open the coin sweep with the strategy, timeframe, window and that row's values filled in, which is how you find out whether a plateau belongs to the strategy or to one coin.

An opened study keeps its id in the address, so a reload or a shared link reopens it rather than starting another.

Memory and disk

A study is held to the machine's free memory and free disk, with the same pause and stop floors as a sweep. A study that stops this way keeps the trials it finished and says so.