Statistical and analytical consulting
Our scientists are experts in a wide variety of statistical methodologies, analytical methodologies, and subject matter areas. We are big proponents of Bayesian modeling; expert opinion matters in every analysis. At the beginning of any relationship, we work closely with you to incorporate your hard-won and often subjective expertise into our models. In other words, we’re data-driven, but we know that experience is invaluable.
For more information, please contact us at firstname.lastname@example.org.
- Multivariate / cross-sectional data analysis with both long (many observations) and wide (many predictors) tables
- Time series analysis and prediction, including automated model selection and dynamic prediction
- Design of (distributed) computer experiments and A/B testing
- Nonparametric analysis, e.g. distribution finding and fitting, error prediction, nonparametric hypothesis testing
- Building and fitting large graph-based models (Bayesian DAGs) of multivariate or time series data
We strongly believe in the validity of Occam’s razor: when two models have similar predictive accuracy, we accept the simpler explanation. With this in mind, we begin any analysis with the simplest models, moving toward models of increasing complexity only when simpler models have unacceptably low predictive power.
All of our machine learning is done using Bayesian methods. When we use complicated models, you can rest assured that we will always provide error estimates and detailed summaries of model criticism; we will never present you with a black-box model and say “it just works, but we don’t know why!”
Our services include
- Traditional supervised learning methodologies (GLM, support vector machines, decision trees and ensembling methods)
- Density estimation and nonparametric modeling for question-answering and what-if sceenarios
- Unsupervised algorithms, e.g., clustering
- Supervised deep learning methodologies for categorical or real-valued data, including image and time-series classification and prediction problems
- Unsupervised and self-supervised deep learning for dimensionality reduction and synthetic data generation
Our scientists have a strong mathematical background and are well-versed in many analytical methods; we can help you develop software to solve high-dimensional optimization or control problems, for example.
Mechanistic and agent-based modeling
We separate ourselves from the competition by asking and answering why things happen and not just settling for prediction. If you understand the mechanism generating observed phenomena, you are better positioned to prepare for situations for which there is no data, situations that haven’t happened yet but may be very costly if not prepared for. One of our specialties is creating mechanistic models of complicated systems that can be interrogated to understand observed behaviors and predict future scenarios.
Examples of such models that we have built include:
- Real-options valuation model of a possible acquisition target for a federal contractor
- Financial market microstructure-based model to understand tail risk and anomalous volatility
- Game-theoretic election interference model to find optimal strategies for countering election meddling