How many parameters is too many?

The count matters less than what you did with it. Ten fixed parameters cost nothing. Three you searched over cost a great deal.

A parameter is only expensive if you optimised it

A setting you chose for a reason and never revisited is an assumption. A setting you swept across a range and kept the best of is a search, and a search finds the best fit to the history whether or not any edge exists.

So the question to ask of a strategy is not how many inputs it has. It is how many combinations you evaluated before choosing this one, because that number is what decides how impressed you should be by the winner.

The number nobody writes down

Three parameters swept across ten values each is a thousand combinations. The best of a thousand random strategies looks excellent, and the optimiser will hand it to you with a smooth equity curve and no warning attached.

Write the count down before you start. It is the single most useful thing to know when reading your own result, and it is impossible to reconstruct afterwards.

Look at the neighbourhood, not the peak

A parameter setting that works while its neighbours fail is a spike, and spikes do not survive contact with a market that has moved slightly. A setting surrounded by other settings that also work is a plateau, and a plateau is what a real effect looks like when it is measured imprecisely.

Prefer the middle of a broad plateau to the top of a narrow peak, even when the peak tests better. You are choosing the result that is most likely to still be there next quarter.

What to do instead of adding another input

Fix every parameter you can justify from the idea rather than from the results. Sweep only what you genuinely do not know. Then test the chosen setting on a window you have never looked at, because that is the only measurement your search has not already spent.

Questions

Is there a maximum number of parameters?

No useful one. What constrains you is how much independent data you have and how many combinations you evaluated against it. Two heavily optimised parameters can overfit a short history more thoroughly than ten fixed ones.

Does walk-forward optimisation solve overfitting?

It measures whether the process of re-tuning would have worked through time, which is a much better question than whether one fixed setting fits. It does not make an exhaustive search safe, and it can itself be repeated until it passes.

What if the strategy only works with very specific settings?

Treat that as the result. A narrow peak is evidence that what you found is a feature of this particular history, which is exactly the finding a backtest is for.

Sonar Sciences is a research and validation platform, not an investment adviser. Nothing on this page is investment advice or a recommendation to trade any instrument or strategy. Platform names are trademarks of their respective owners and we are not affiliated with them. Trading leveraged instruments involves substantial risk of loss.