The appropriateness of using a normal distribution for Volkswagen's log returns can be evaluated by computing skewness, kurtosis, and performing a Jarque-Bera test. Skewness measures asymmetry, with a value near zero indicating symmetry typical of normal distributions. Kurtosis measures tailedness, with a value of 3 indicating normality. Significant deviation from these values suggests non-normality. The Jarque-Bera test, which combines skewness and kurtosis to test the normal hypothesis, if returning a low p-value, indicates that log returns are not normally distributed. Thus, if skewness and kurtosis deviate significantly from 0 and 3, respectively, and the Jarque-Bera test suggests rejection of normality, a normal distribution model is inappropriate .

Assuming time independence and stationarity in financial time series, like the DAX index, suggests that returns are consistent in distribution over time, without dependence on past outcomes. This assumption simplifies modeling as it allows use of linear models and time-invariant coefficients. However, financial markets often show volatility clustering and trend components, violating these assumptions. Ignoring such features may lead to inaccurate predictions and misinterpretations of market dynamics, suggesting the use of these assumptions with caution or the introduction of models accounting for time-varying volatility like ARCH or GARCH processes .

The covariance of the least squares estimator \( \hat{\beta} \) is given by \( \operatorname{Cov}(\hat{\beta}) = \sigma^2(X^TX)^{-1} \), conditioned on the design matrix \( X \). To derive an unbiased estimator for the variance, \( \hat{\sigma}^2 \), we use \( \hat{\sigma}^2 = \frac{1}{n-k-1} \hat{\epsilon}^T \hat{\epsilon} \), where \( \hat{\epsilon} = y - X\hat{\beta} \) denotes the residuals. This expression leads to an unbiased estimator for \( \operatorname{Var}(\hat{\beta}) = \hat{\sigma}^2 (X^TX)^{-1} \), assuming normality of errors .

For the non-linear model \( f(z_i, \beta, \epsilon_i) = \exp(-\beta_1 z_{i1} \exp(-\beta_2 z_{i2}) \epsilon_i) \), to linearize it, one can take the logarithm of both sides, yielding \( \log(y_i) = -\beta_1 z_{i1} - \beta_2 z_{i2} + \log(\epsilon_i) \). This transformation aligns with a linear model \( Y' = Z\beta + \epsilon' \) where \( Y' = \log(y_i) \) and \( Z = [-z_{i1}, -z_{i2}] \) with \( \epsilon' = \log(\epsilon_i) \). This allows us to use linear regression methods to estimate \( \beta \) from the model .

The approach proposed for estimating the deterministic drift of \( \log(S_t) \) in the DAX index is to model \( S_t = \beta_0 + \beta_1 t + \epsilon_t \) and use least squares estimation for \( (\beta_0, \beta_1)^T \). This assumes a linear trend in stock prices. However, this approach may not fully capture the complexity of financial time series which can exhibit non-linear trends, volatility clustering, and leptokurtosis. Thus, while it may provide some insights into a general trend, failing to account for these nuances can lead to biased or oversimplified models for forecasting .

To demonstrate that \( \hat{\beta} \) is an unbiased estimator, we rely on the key property that the expected value of the estimator equals the parameter being estimated. In the classical linear model \( y = X\beta + \epsilon \), the least squares estimator \( \hat{\beta} \) is given by \( (X^TX)^{-1}X^Ty \). Taking expectations, \( E[\hat{\beta}] = (X^TX)^{-1}X^TE[y] \). Since \( E[y] = X\beta \), we get \( E[\hat{\beta}] = (X^TX)^{-1}X^TX\beta = \beta \), confirming that \( \hat{\beta} \) is an unbiased estimator for \( \beta \).

Non-linear regression functions like \( f(z_i, \beta, \epsilon_i) = \beta_1 z_i^{\beta_2} \epsilon_i \) can be transformed into linear models by applying logarithmic transformation to linearize the exponents. Taking log on both sides, we have \( \log(y_i) = \log(\beta_1) + \beta_2 \log(z_i) + \log(\epsilon_i) \). This describes a linear relationship between \( \log(y_i) \) and \( \log(z_i) \) with coefficients \( \log(\beta_1) \) and \( \beta_2 \). The transformed model allows the estimation of \( \beta \) using least squares on the transformed variables .

The covariance \( \text{Cov}(\hat{\beta}) = \sigma^2(X^TX)^{-1} \) is calculated under the assumption that the errors \( \epsilon \) are homoscedastic and uncorrelated, meaning that \( \,\text{Cov}(\epsilon) = \sigma^2I \). Independence among errors implies no correlation between errors for different observations, which ensures the validity of this covariance matrix. If the independence condition is violated, resulting heteroscedasticity or autocorrelation would require alternative estimators or adjustments such as using heteroscedasticity-robust standard errors to provide unbiased variance estimates .

A QQ-plot compares the quantiles of the log returns of a stock against those of a fitted normal distribution. Points lying on a 45-degree reference line indicate that the data's distribution aligns with a normal distribution. Deviations from the line suggest disparities in distribution; points above the line in tails indicate heavy tails (leptokurtosis), while deviations in the center indicate skewness or other central distributional differences. A systematic deviation indicates that the data may not follow a normal distribution, suggesting the need for alternative models or distributions to better capture the data characteristics .

To compare the log returns of BMW and Volkswagen stocks, first calculate the log returns for each over the given period. Next, visualize these returns on a time series plot for a direct comparison of their volatility and trends. A scatter plot of BMW against Volkswagen returns can reveal any linear correlation, where clustering along a line indicates correlation. Statistical analysis such as calculating Pearson or Spearman correlation coefficients offers insight into the strength and type of relationship. Further, regression analysis can evaluate how changes in one stock's returns are associated with changes in the other's .