Optimisation is the process of choosing a strategy's parameters. It is also the fastest way to produce a strategy that works beautifully on history and fails immediately in the market.
The methods below differ in how efficiently they search. None of them protects you from the central danger, which is that the parameter set with the best backtest is usually the one most fitted to noise.
Adjusting the numbers a strategy depends on: lookback lengths, thresholds, stop distances. The goal is not the highest score. It is finding values that are not unusually sensitive, so that being slightly wrong about the market does not break the strategy.
Testing every combination on a defined grid of parameter values. Exhaustive, easy to reason about, and expensive: cost grows exponentially with the number of parameters. Practical for two or three, unworkable for ten.
Sampling parameter combinations at random instead of exhaustively. Counter-intuitively it usually beats grid search when many parameters barely matter, because it spends its budget exploring the few that do rather than mapping the ones that do not.
Building a probabilistic model of how parameters map to results, then choosing the next test where improvement is most likely. Far more sample-efficient than grid or random search, which matters when each backtest is slow.
Treating parameter sets as a population: keep the best performers, combine them, introduce random mutation, repeat. Good at escaping local optima in awkward search spaces, and prone to overfitting if run for too many generations against one dataset.
Deciding which inputs a strategy should use at all. Adding indicators almost always improves a backtest and often worsens live results, because each new input is another chance to fit noise. Removing inputs is usually the higher-value work.
Optimising for several goals at once, typically return against drawdown. Instead of a single best answer you get a frontier of trade-offs, and you choose where on it you want to sit. This is more honest than collapsing everything into one score.
Deliberately choosing parameters that perform acceptably across many conditions rather than optimally in one. In practice this means preferring the middle of a broad plateau of decent results over a narrow peak, because the peak is usually an artefact.
Letting values respond to conditions instead of staying fixed, such as sizing stops by current ATR rather than a set percentage. It helps a strategy survive regime changes, at the cost of adding complexity that itself needs validating.
Running several strategies together so their weaknesses do not coincide. The benefit comes from low correlation between them: two profitable strategies that lose at the same time diversify nothing.
Aggregating signals from multiple models, by voting or weighting, into one decision. Ensembles are typically steadier than their best member because individual errors partly cancel, which is the same reason they rarely post the highest headline return.
The Strategy Studio and backtester make the tuning loop fast. Speed is a double-edged tool here, because the faster you can search, the easier it is to overfit, so the discipline still has to come from you.
No screens to babysit: define the conditions and the bot does the rest.
