The Claude Quant PlaybookDownload ebookDownload ebook

Mean squared error: bias² plus variance

intermediateStatistics10 of 10

The average squared error of an estimate splits into two parts: systematic error (bias, squared) and noise (variance). Accepting a little bias can remove a lot of noise.

Build a trading system with Claude.Download ebook

01The formula

an estimator: a rule that turns data into a guess
the true value being estimated, here the mean μ
how far the estimator is off on average
how much the estimator moves from sample to sample
shrinkage: how far the sample mean is pulled towards zero
the sample mean
the true mean
the volatility of one observation and the number of observations

02See it move

03Where it is used

04Why traders use it

Unbiased estimates of mean returns are so noisy that they make poor inputs. Shrinking them is one of the most reliable improvements in practice, and this split shows why.

05Worked example

With a true mean of 0.05%, volatility 1.0% and 250 days, shrinking the sample mean by 30% gives an MSE of 21.8 bp², made of bias² 2.3 and variance 19.6. The best shrinkage here is 62%.

06How it is built

  1. Add and subtract : .
  2. Square and take expectations; the cross term has mean zero, leaving .
  3. For : the bias is and the variance .
  4. Minimising over gives : shrink more when the data are noisy.

07The common mistake

Insisting on unbiased estimators. An unbiased estimate with huge variance is usually further from the truth than a slightly biased, much steadier one.

08In Python and LaTeX

mu, sigma, n, s = 5.0, 100.0, 250, 0.3   # in basis points
bias2 = (s*mu)**2
var = (1 - s)**2 * sigma**2 / n
print(round(bias2, 2), round(var, 2), round(bias2 + var, 2))
\operatorname{MSE}(\hat\theta) = E\big[(\hat\theta - \theta)^2\big] = \operatorname{Bias}^2 + \operatorname{Var}, \qquad \hat\theta = (1 - s)\bar{x}:\ \ \operatorname{MSE} = s^2\mu^2 + (1 - s)^2\frac{\sigma^2}{n}

09Where it fits

Rather not code the maths yourself? Describe your trading idea and let AI build the bot.

Build a bot with AgenKit

The worked example and the model run the same code, so every number matches. Models simplify markets on purpose. This is education, not financial advice.