Digital Twin & Simulations
Once a causal graph has been discovered and validated, RootCause creates a Digital Twin: a live, data-driven model of your system. The twin acts as a sandbox where interventions, counterfactuals, and optimizations can be tested out, and their impacts evaluated, before being applied in real life.
Definition & Purpose
The Digital Twin is the execution layer of RootCause. It translates causal structure into decision support by:
Simulating the effects of interventions in a controlled environment
Providing explainable reasoning behind KPI changes
Balancing trade-offs across multiple outcomes
This makes it possible to move beyond descriptive analytics and into prescriptive, causally sound decision-making.
How It Works
Baseline World – The twin samples outcomes from the learned causal model.
Intervention – One or more variables are modified (hard values, relative changes, or segment-specific).
Propagation – Effects flow through the causal graph, updating downstream nodes according to their dependencies.
Simulation Runs – Monte Carlo sampling produces distributions of possible futures.
Comparison – Baseline and intervention scenarios are compared, with uncertainty intervals provided.
Advanced Capabilities
Bayesian Foundations – The twin runs on a causal Bayesian network whose conditional distributions are learned with Bayesian additive regression tree (BART) ensembles and sampled via Monte Carlo. In temporal causal discovery, Gaussian-process decomposition separates trend and seasonal structure before dependencies are tested.
Ontology Integration – Variable dependencies are inferred from the ontology, ensuring simulations respect real-world structure.
Scalability – A staged filter cascade prunes the pairwise search space before expensive independence tests run, keeping discovery tractable on high-dimensional datasets.
Segment Analysis – Simulations can run across sub-populations (regions, customer cohorts, product lines) to uncover heterogeneous effects.
Optimization – The system can recommend levers that maximize or minimize a target while accounting for secondary impacts.
Natural Language Interface – Simulations can be configured via structured UI or plain-language queries.
Deployment & Performance
Enterprise-Ready – Runs self-hosted for sensitive environments, with optional cloud execution for evaluation.
Built for Volume – Memory requirements are estimated up front before discovery starts, panel data is sharded across entities, and the heavy computation runs in native Rust and C++ kernels, including tiled distance-correlation computation for wide datasets.
Efficiency – Designed to operate in high-dimensional feature spaces where traditional causal inference becomes infeasible.
Types of Simulations
Predictions – Estimate the outcome of a specified input, with uncertainty.
Interventions – Test the effect of changing a driver variable.
Optimizations – Automatically search for the best intervention to reach a desired outcome.
Best Action – Find the minimal change that reaches a target outcome.
Explanations – Identify which drivers most influenced a KPI shift.
Root Cause Analysis & Anomaly Scan – Diagnose why a value came out anomalous.
See Simulation Types for a practical guide to each simulation.
Last updated