Maximum Likelihood Estimation Finance: Fitting Models to Data

Key Takeaways

• MLE maximizes the log-likelihood, which converts products of densities into a sum that is numerically stable and easy to differentiate.
• For a normal model, MLE recovers the sample mean and variance; for a Student-t, it requires numerical optimization that yields a degrees-of-freedom near 4.5 on equity data.
• Running MLE from a single starting value risks converging to a local maximum in non-convex models like GARCH or mixture distributions.
• Adding parameters always improves in-sample log-likelihood; AIC, BIC, or out-of-sample tests are needed to prevent overfitting.

What It Is

Given a probability model with parameters theta and a dataset of observations, the likelihood function is the joint probability of the data treated as a function of theta. MLE picks the theta that maximizes this function. In finance it is the default way to fit volatility models, return-distribution models, and option-pricing calibrations.

The method goes back to R.A. Fisher in the 1920s and is a mainstay of modern econometrics. Its appeal is practical: under mild regularity conditions, MLE estimators are consistent (they converge to the true parameter as sample size grows), asymptotically efficient (they reach the Cramer-Rao lower bound on variance), and asymptotically normal (you can build standard errors and confidence intervals).

The Intuition

Suppose you have 500 daily returns and you want to fit a Student-t distribution to them. There are many possible combinations of mean, standard deviation, and degrees of freedom. MLE asks: which combination would have produced a dataset that looks most like the one I actually observed?

The answer is found by writing down the probability density at each observed data point given the candidate parameters, multiplying them (assuming independence), and picking the parameters that make the product largest. In practice you maximize the logarithm of this product, which turns products into sums and stays numerically stable.

How It Works

For an i.i.d. sample of observations x_1 through x_N with density f(x | theta), the log-likelihood is:

logL(theta) = sum_{i=1}^{N} log f( x_i | theta )

The MLE estimate is:

theta_hat = argmax_theta  logL(theta)

Solving this typically requires numerical optimization: gradient ascent, Newton-Raphson, or BFGS. Most statistical packages handle it automatically.

Once you have theta_hat, the observed Fisher information is the negative second derivative of the log-likelihood evaluated at theta_hat:

I(theta_hat) = - d^2 logL / d theta^2   (evaluated at theta_hat)

The inverse of this matrix gives the asymptotic covariance of the estimator, from which standard errors and Wald tests follow.

Typical finance applications:

• GARCH fitting. The conditional likelihood of returns given a GARCH(1,1) variance process is maximized to recover omega, alpha, and beta.
• Distribution fitting for VaR. Student-t degrees of freedom, generalized Pareto tail index, skewed-t parameters, all estimated by MLE on historical returns.
• Option-pricing calibration. Heston stochastic-volatility or SVI surfaces are fit by minimizing a likelihood or pricing error across observed option quotes.
• Regression with non-Gaussian errors. When errors are known to have fat tails or heteroskedasticity, MLE with an appropriate density replaces ordinary least squares.

Worked Example

Fit a normal distribution to 1,000 daily returns with sample mean 0.05 percent and sample standard deviation 1.2 percent.

For a normal density f(r | mu, sigma), the log-likelihood is:

logL(mu, sigma) = -(N/2) log(2 pi)  -  N log(sigma)  -  (1 / (2 sigma^2)) sum (r_i - mu)^2

Taking derivatives and setting them to zero:

mu_hat    = sample mean           = 0.0005
sigma_hat = sample std deviation  = 0.012

So for a normal model, MLE gives back the familiar sample moments. The value of MLE becomes clear when you swap the normal for a Student-t with three parameters, where closed forms no longer exist and numerical optimization is required. The same log-likelihood framework handles both cases.

For a Student-t fit on typical equity data, the numerical maximum often lands near nu_hat = 4.5, mu_hat close to zero, and a scale parameter slightly below the sample standard deviation because fat tails inflate the raw sigma.

Common Mistakes

• Assuming independence when data are autocorrelated. Plain MLE on daily returns assumes observations are i.i.d. If you include returns that are serially correlated (intraday bars, short-horizon index returns), standard errors are understated. Use quasi-MLE with Newey-West or HAC covariance corrections.

• Finding a local maximum instead of the global one. Non-convex likelihoods (common in GARCH, stochastic volatility, mixture models) can have multiple peaks. A single run from one starting value can converge to a bad local maximum. Always restart from several initial values and check convergence.

• Confusing MLE with model selection. A fitted likelihood is not a score for comparing different families of models. Use AIC, BIC, or out-of-sample tests. Two models fit with MLE to the same data can both maximize their own likelihoods without being comparable.

• Overfitting by adding parameters. Every extra parameter improves the log-likelihood mechanically. Penalized-likelihood (AIC, BIC), cross-validation, or Bayesian priors are the corrections. Raw MLE on a high-dimensional model will overfit happily.

Frequently Asked Questions

Q: What is maximum likelihood estimation in finance in simple terms? It is the method of choosing model parameters by asking which values would have made the observed data most probable, then solving for those values numerically or analytically.

Q: How does maximum likelihood estimation in finance affect investment decisions? It provides the fitted parameters for GARCH volatility forecasts, fat-tailed VaR models, and option-pricing calibrations, all of which directly determine position sizing, hedging ratios, and risk budgets.

Q: What is a real-world example of maximum likelihood estimation in finance? Fitting a Student-t distribution to ten years of daily S&P 500 returns by MLE returns degrees-of-freedom near 4.5, confirming heavy tails, while a Gaussian MLE fit would understate 99 percent VaR by 30 to 50 percent.

Q: How can investors use maximum likelihood estimation in finance? Use MLE to fit realistic return distributions to your specific asset class and time period, compare competing model families using AIC or BIC, and always restart the optimizer from multiple initial values when fitting nonlinear models like GARCH.

Q: How is maximum likelihood estimation in finance different from Bayesian inference? MLE produces a single point estimate that maximizes the likelihood, treating parameters as fixed unknowns. Bayesian inference produces a full posterior distribution over parameters, incorporating prior beliefs and providing credible intervals that honestly reflect uncertainty in small samples.

Sources

1. Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Elements of Statistical Learning, 2nd edition. Springer. https://hastie.su.domains/ElemStatLearn/
2. Bollerslev, T. (1986). "Generalized Autoregressive Conditional Heteroskedasticity." Journal of Econometrics, 31(3), 307-327. https://doi.org/10.1016/0304-4076(86)90063-1
3. Greene, W.H. (2018). Econometric Analysis, 8th edition. Pearson. https://www.pearson.com/en-us/subject-catalog/p/econometric-analysis/P200000006213
4. CFA Institute. Quantitative Methods curriculum readings. https://www.cfainstitute.org/programs/cfa-program/candidate-resources

Disclaimer

This article is educational content only and is not financial advice. Nothing here is a recommendation to buy, sell, or hold any security. Consult a licensed advisor before making investment decisions.