Solutions

Data have little meaning without proper statistical analysis—it is statistics and data science that unlock the value within data.

Doing data science and statistics properly is a significant responsibility that demands specialised expertise. We employ methods that are as complex as necessary, with a strong emphasis on communicating actionable results. Our commitment to quality, transparency, and innovation ensures that our tailored solutions deliver meaningful impact.

Answering Business Questions with Data Science

We support your company in identifying business-relevant questions and answering them with data science. Throughout this process, we prioritise confidentiality and the seamless integration of results into your real-world operations. By applying state-of-the-art machine learning techniques for forecasting and estimating forecast precision, we help you make informed decisions and optimise your operations.

Data-Driven Decisions for Campaigns and Policies

We support regional, cantonal and national administrations and governments in taking data-driven actions, from the design of campaigns to the implementation of policies. We advise on and deliver custom software solutions for data collection, data integration, data analysis, and modeling.

Support for Data-Driven Research

We collaborate with researchers by supporting them in all stages of their scientific project: the formulation of research questions, the design of the study, the choice and implementation of a sound statistical analysis, the write-up of scientific publications and the navigation through the peer-review process. Our network of contacts with specialised knowledge includes numerous leading experts both from the private industry and academia.

Teaching Data Science & Statistics

We develop tailored upskilling courses, presentations, lectures and training sessions for universities, private companies and administrations. Our teaching can be held online or in presence according to your needs. We are at the forefront of fast-paced science by lecturing at several top-ranked universities such as ETH Zurich and the University of Oxford.

Products we offer

Here are the products that we offer.

We forecast/predict future events such as daily customer spending in all your shops with the goal to best allocate your sales employees. The forecast’s precision is assessed by estimating how reliable the forecasts are. We estimate how well the forecasting model would perform on new, unseen data to judge how well the model is applicable in daily business.

We estimate the effect of air temperature on ice cream sales or the effect of a medical treatment on a health outcome - whatever effect you might be interested in that can be investigated with your data. We assess the estimate’s precision by estimating how reliable the estimations are. If possible, we apply causal inference to best distinguish causal relationships from spurious correlations, in order for you to take data-driven decisions.

If you already use a statistical or machine learning model, we make an in-depth assessment of the suitability of the model: we check its mathematical assumptions and how likely they are verified, whether the model is suitable to answer the posed questions and whether the model is applicable in a given real-world setting. We describe and/or implement the steps to improve your model.

We design your study along the following steps: defining clear research questions, selecting an appropriate study design including a suitable randomisation scheme, determining the needed sample size, and defining a statistical analysis plan that fits the research questions and the data that will be collected.

We design comprehensive online and on paper questionnaires with questions targeting to collect the right data to answer your research questions. We distribute the questionnaires, and collect, digitalise and anonymise the answers.

We design and implement user-friendly interfaces in your corporate design for efficient data collection, handling and visualisation. We create interactive and visually appealing dashboards to monitor key metrics (for example of a statistical model). We also develop APIs for live implementation of statistical and machine learning models on a server.

We set up templates that, based on specific parameters, become customised reports for repeated use, for example to summarize your statistical model’s performance every month using the newest data input. This approach ensures consistency and accuracy in every report, saving you a lot of time.

We plan and conduct the statistical analysis for your scientific project, write-up the statistical part and interpretation in your paper and help you navigate through the whole peer-review process. We ensure strict adherence to the highest scientific standards throughout this process.

Methods we use

We are experts in the following statistical methods, which we use to either build predictive models (for forecasts) or to estimate (causal) effects and their uncertainty.

We are experienced in classical statistical methods for randomised controlled trials and observational data in medical research. These include comparison of treatment and control groups, survival analysis, methods to account for repeated measures of study subjects, and uncertainty estimation.

Would the number of accidents at complex crossroads decrease if we installed highly visible stop signs? Would a stop sign campaign for complex crossroads be worthwhile?

To answer these questions, we need to estimate the effect of the “installing stop signs” intervention on the number of accidents. A (simple) estimator for such a measure can lead to a spurious correlation that is not a causal relationship. If we actually installed stop signs, the change in the number of accidents might deviate significantly from our model’s predictions. The problem are so-called “confounders”, which can distort our estimator.

When making decisions or giving recommendations to policymakers, we need solid estimates for quantities that cannot be observed directly: It is not possible to measure the number of accidents at a crossroads under the same conditions both with and without a stop sign. An experiment with randomized stop signs would allow us to estimate the causal relationship between “installing stop signs” and the number of accidents.

Even when such an experiment cannot be conducted, the fascinating field of causal inference provides tools and methods to distinguish spurious correlations from causal relationships. We have strong expertise in these methods and are experienced in formulating your questions within the framework of causal inference. We know the classical causal inference tools such as directed acyclic graphs (DAGs) and the potential outcomes framework, but also more advanced tools that incorporate machine learning.

High-dimensional statistics focuses on analysing data sets with a large number of variables relative to the number of observations. As traditional methods may not perform well in high dimensions, this field has become increasingly important due to the rise of big data and complex data structures in various disciplines, such as genomics and finance.

Many high-dimensional data sets are sparse, meaning most variables have little or no relationship to the response. Techniques like LASSO help in selecting the most relevant variables. Methods such as PCA reduce the number of dimensions while preserving important data structures and relationships. We are experts in these well established as well as new methods for hypothesis testing and constructing confidence intervals to get valid statistical inference in high-dimensional settings.

Machine learning describes algorithms that enable computers to learn from data and make predictions. We are experienced in applying some of the most relevant methods in supervised learning (support vector machines, neural networks and random forests) and unsupervised learning (e.g., k-means clustering and principal component analysis).

We are also experienced in conformal prediction (CP), a framework to quantify the uncertainty of any prediction, even if it is obtained through non- or semi-parametric methods like machine learning, for which this was not possible until recently. Also, we are experts in the most recent methods combining the predictive power of machine learning with the principled framework of causality, which allows for valid statistical inference.

Mixed models are statistical models that include both fixed and random effects. These models are useful in various disciplines, including the biology and social sciences. They are especially useful if the data contains repeated measurements of the same units (as in longitudinal studies) or measurements on clusters of related units. Mixed models thus do not assume independent observations, which is very often the case in practice. We are experts in the application and assessment of mixed models as we have used them during many years and in many projects.

Transformation models are a class of models for generalised regression. They present an advantageous approach to modelling when one wants to find a sweet spot between flexibility and interpretability. Their versatility allows their application to any type of tabular data. This class of methods is currently being developed and researched extensively by several research groups in Europe.

Robust methods can be applied on top of many other methods, with the goal to make them more robust to model misspecification, outliers and violations of model assumptions.

Workflow we follow

Below is an overview of the key steps in our workflow. Depending on the specific project, only some of the steps are performed by us.