Degrees of freedom (strategy parameters versus trades)
In strategy research, degrees of freedom are the choices a rule is allowed to fit to the data: parameters, filters, thresholds and discretionary design decisions. Each one lets the rule adapt to the sample, so the more of them relative to the number of trades, the less the backtest says about new data.
Senzoukria · Glossary · Updated September 2026
Counting what the rule can fit
A moving-average crossover has two lengths. Add a session filter, a volatility threshold and a stop distance and there are five tunable quantities. If each takes five candidate values, the grid holds 5⁵ = 3,125 combinations, compared with 25 for the two lengths alone. The number of configurations the selection can choose from is a direct measure of how much room the rule has to fit noise.
The hidden degrees of freedom matter as much as the declared ones: the instrument, the timeframe, the bar type, the trading window and each rule added after looking at results. They are rarely counted and always used.
Trades per degree of freedom
Practitioners often ask for many trades per free parameter before trusting a fit. There is no exact threshold, because trades are not independent and parameters do not all have the same influence, but the direction is reliable: a four-parameter rule with 40 trades has fitted about as much as it has measured. Out-of-sample testing, parameter stability checks and a deflated Sharpe that knows the grid size are the tools that turn this intuition into numbers.
The statistical sense: N or N − 1
In statistics, degrees of freedom also refers to the denominator of a sample variance. Dividing the squared deviations by N describes the sample you have; dividing by N − 1 gives an unbiased estimate of the population variance. On small samples the difference is visible: with 10 trades, the N − 1 standard deviation is √(10/9) ≈ 1.054 times the N version, so a Sharpe ratio differs by about 5% depending on the convention.
In Senzoukria
The sweep and the walk-forward of the automatic backtest cap the grid at 400 combinations and display the number of combinations times slices before a run; the button that explores every readable parameter of a script sets up to five axes around its defaults. The Gauntlet's deflated Sharpe uses the full grid size as the number of trials. Both conventions of the standard deviation appear, each documented in the source: the Replay statistics use the population version (divide by N), while the Gauntlet reproduces the research backtester with the sample version (N − 1) and a square-root-of-252 scaling of daily results.
Related
In the same section
- Delayed data
- Deflated Sharpe
- Delta
- Dealer positioning
- Delta at extremes
- Databento
- Delta bars
- Data snooping bias
This page in other languages
Frequently asked questions
- How many parameters can a strategy have?
- As few as the idea requires. Every additional parameter must earn its place by improving results out of sample, not in sample. If removing a parameter barely changes out-of-sample performance, it was mostly fitting the training data.
- Why do two tools give different Sharpe ratios on the same trades?
- One common reason is the variance convention: dividing by N or by N − 1. Another is annualization. On large samples the first difference is negligible; on a dozen trades it is several percent.