Can the Böögg predict Zurich's summer?

An Analysis Through Tradition and Data Science

Introduction

Zurich's Sechseläuten festival is one of the city's most cherished traditions, blending history, community, and celebration to mark the transition from winter to summer. At the heart of this festival is the burning of the Böögg, a snowman effigy symbolising winter, whose fiery demise is believed to predict the nature of the upcoming summer. For decades, spectators have watched with anticipation as flames climb the pyre, with the time it takes to reach the Böögg's head becoming a folk-based summer forecast. But how accurate is this tradition? Can the Böögg truly foretell Zurich's summer, or are there more dominant factors at play? In this article, we'll assess using data science whether the Böögg's burning time has any real predictive power or if modern meteorological insights provide a clearer picture of what Zurich can expect each summer.

Why Data Science?

The Böögg's burning time has long been a source of intrigue, but how well does this tradition hold up against the rigor of scientific analysis? Data science offers a powerful way to test the validity of this folklore by examining whether there is a meaningful correlation between the Böögg's burning times and Zurich's summer weather conditions. By combining historical records of burning times with meteorological data, we can uncover patterns, trends, or anomalies that might validate or challenge the Böögg's predictive power. Key data sources for such an analysis include archives of Böögg burning times, Zurich's detailed weather records, and broader climatic factors such as global temperature anomalies or seasonal atmospheric trends. This approach not only brings clarity to a cultural tradition but also demonstrates how modern tools can bridge the gap between folklore and science, providing a fresh perspective on long-held beliefs.

Data Gathering and Initial Analysis

To test the Böögg's ability to predict Zurich's summer weather, data from two primary sources was gathered: historical records of Böögg's burning times and local weather data. With this dataset in hand let's evaluate whether the Böögg's burning time (seconds - see right hand axis below - black line) - a single feature - shows any meaningful correlation with Zurich's summer temperatures (target variable on the left hand axis - blue line).

Let's perform a linear regression to analyse the relationship between the datapoints of the two lines:

Even though the RMSE is relatively low, it's visually obvious that the correlation (see R²) is very weak between the burning time and the average summer temperature.

Root Mean Squared Error (RMSE): 1.35
R-squared (R²): 0.07

Note: The regression has been calculated using a mean centered feature variable increasing the interpretability of above equation. Following the centering, the intercept represents the predicted value of y when the predictor (x) is at its mean value (zero after centering). Without centering we would obtain a highly negative intercept to compensate for the year going into the equation. Regardless whether the features are centered, scaled or used as-is, the errors respectively R² stay constant. The above equation shows that the duration almost has no impact on the target temperature as the slope is rather flat. 

\[  Temp = 22.17 + 0.0005 \cdot duration \]

Let's have a look at the year as an alternative feature instead of the burning duration of the Böögg:

Root Mean Squared Error (RMSE): 1.24
R-squared (R²): 0.21

\[  Temp = 22.17 + 0.07 \cdot year \]

Above equation indicates that for each additional year (relative to the mean year), the temperature is predicted to increase by approximately 0.07 degrees Celsius The mean absolute error has slightly decreased which again looks like a small average prediction error. The R² of 0.21 still does not explain much of the variability and there might be other relevant features not included in the model having an impact on the target variable.

Having analysed two features isolated, let's turn to supervised learning and analyse them together using a decision tree which evaluate possible split points for the features to minimise the variance in the target variable.

Above graph indicates that the year is a significant feature for predicting summer temperatures. For years earlier than 2015, shorter burning durations correlate with higher predicted temperatures, while longer burning durations correlate with lower temperatures. For more recent years (after 2014.5), burning duration still impacts the prediction but with slightly different thresholds and outcomes.

Why would we want to use a decision tree? Decision trees are easy to understand and interpret and they can also model non-linear relationships. The tree structure allows you to visualise the decision-making process, making it clear how the model is making predictions based on the features. This can be particularly useful when you need to explain the model's decisions to stakeholders.

Let's in addition review the importance of the single features as a next step using random forest. Note: A random forest is an ensemble learning method used for both classification and regression tasks. It operates by constructing multiple decision trees during training and outputting the mode of the classes (classification) or mean prediction (regression) of the individual trees.

# create a regressor object 
regressor = RandomForestRegressor(random_state = 42, max_depth=2)  
  
# fit the regressor with X and Y data 
regressor.fit(x, y)

feature_importance = pd.Series(regressor.feature_importances_, index=x.columns)

Running above script results in the following feature importance: year 67%, duration 33%
This underlines our observations that the year is a better feature to predict the summer average temperatures than the burning duration of the Böögg.

For the year feature let's look in addition at regularisation to add a penalty to the model's complexity. Regularisation reduces overfitting by penalising large coefficients and encourage simpler models that generalise better.

There are several options like Lasso, Ridge or HuberRegressor - we pick the last one and use Grid Search CV to evaluate the best hyperparameter combination for the model:

Root Mean Squared Error (RMSE): 1.29
R-squared (R²): 0.14

The regularised model which concludes this section shows even worse results. It outlines that none of the approaches could provide a satisfying model and that the underlying features link to the target variable may be too weak after all.

Findings

The initial linear regression analysis revealed no significant causal relationship between the Böögg's burning time and Zurich's summer temperatures. Even after applying regularisation techniques such as ridge and lasso regression to reduce overfitting, the model failed to identify a meaningful link between burning time and summer weather patterns. This suggested that either the Böögg's burning time is not a strong predictor of summer temperatures or that additional factors need to be considered.

As a next step, I introduced the feature year into the dataset, hypothesising that more recent years might have hotter summers due to global warming trends. While this added some predictive power to the model, the Böögg's burning time remained an insignificant feature, further weakening the tradition's claim of predictive accuracy.

To explore the data more comprehensively, I employed decision tree and random forest regression models. These techniques allowed for better handling of non-linear relationships and provided insights into the relative importance of each feature. The results were clear: the year feature emerged as the strongest predictor of Zurich's summer temperatures, highlighting the long-term warming trend likely driven by global climate change. The Böögg's burning time, on the other hand, consistently ranked as one of the weakest predictors, with little to no contribution to the model's accuracy.

Conclusion

The analysis reveals that while the Böögg is a cherished tradition steeped in Zurich's cultural history, its ability to predict summer weather is unsupported by data. Even with advanced analytical techniques, no significant correlation was found between the Böögg's burning time and Zurich's summer temperatures. Instead, broader factors like long-term climate trends, represented by the year feature, proved to be far more reliable indicators of summer weather. This highlights that the Böögg's predictive power is largely symbolic, rooted in folklore rather than scientific evidence. Despite these findings, the tradition of burning the Böögg's remains invaluable as a cultural event that brings people together and marks the transition from winter to summer.

Alberto Desiderio is deeply passionate about data analytics, particularly in the contexts of financial investment, sports, and geospatial data. He thrives on projects that blend these domains, uncovering insights that drive smarter financial decisions, optimise athletic performance, or reveal geographic trends.