How to optimize parameters without losing context
Set up a small search, interpret the Pareto frontier and record what you tested in the Deepwick optimizer.
An optimization can produce hundreds of results and leave you remembering one number: the best. Research also needs a record of what you tried, what you rejected and how performance changes when parameters move.
Deepwick's Backtest Optimizer runs combinations in your browser. This guide sets up a small, documented experiment so you can learn the workflow.
Start with a question and a baseline
Choose a strategy with numeric inputs and keep its original configuration as your baseline. The form lets you select the script, symbol, timeframe, bar count and search ranges. Sign in to access the dashboard optimizer.
Write a hypothesis before running it. For example: “I want to check whether behavior changes gradually as entry length changes.” That asks about sensitivity; it does not assume profitability.
Keep the provider, bars, other inputs and costs constant. An improvement between runs on different historical windows cannot be attributed only to a parameter change.
A bounded search example
For a Donchian strategy, you could investigate entry length from 15 to 25 in steps of 5, and exit length from 5 to 10 in steps of 5. That creates six combinations: three entry values multiplied by two exit values. These are illustrative settings, not parameters recommended for trading.
Write the ranges down before starting and retain the remaining configuration. The combination counter helps identify an unnecessarily large grid. If you expand the ranges after seeing results, record that expansion as another experiment.
Read the Pareto frontier
The optimizer presents results from a Pareto frontier and a net-profit-versus-drawdown plot. The purpose is to compare alternatives offering different trade-offs between those measures. Result cards are sorted by Sharpe within the displayed set.
The cards are not a complete history of every combination. The trial count and visible result count can differ. To investigate a particular neighboring configuration, rerun it explicitly; absence from the list does not mean it was never tested.
A saved backtest's quant report also offers Parameter sensitivity. This sweep uses two numeric inputs, and its map may contain only selected results. A blank cell does not establish either stability or failure.
Close the experiment before extending the search
Record ranges, trial count, baseline settings, trades, drawdown and rejected decisions. Review costs and execution assumptions before comparing a browser result with a saved report.
Repeated selection on the same sample can fit noise; The Probability of Backtest Overfitting studies this selection problem. Similar results from neighboring settings can justify further investigation, but do not certify robustness.
Next, define an evaluation beyond the training window without changing the rule after inspecting the reserved result.