New counterfactual simulator lets marketers query ROI before committing budget
Three-layer architecture combines SCM, Hawkes processes, and LLM agents for causal simulation.
Oransim tackles the problem of prospective marketing attribution. Instead of fitting retrospective models, it lets users simulate interventions before committing budget—like shifting 30% budget from creator A to creator B, or changing publish time by 6 hours. The architecture is deliberately split into three layers: an SCM for clean do()-style interventions over creative-creator-platform-user paths; Hawkes processes for temporal clustering of engagement events (likes, comments, shares); and LLM-driven user archetypes that generate engagement-like signals. These agent signals are consumed as endogenous variables by the SCM, with processing via a modality-agnostic embedding bus. Currently text-only, video/image support is slated for v0.5.
The project is explicit about its relation to existing causal libraries: it differs from DoWhy, EconML, or DoubleML by focusing on simulation given a known graph, not estimation from observational data. The biggest open question is the interface between SCMs and LLM agents—identifiability degrades when an LLM mediates a node, and interventions on prompts don't equal interventions on latent user state. The author welcomes critique on this boundary. The open-source release ships with synthetic data calibrated to public CTR/engagement ranges. Users must bring their own data for real predictions.
- Three-layer stack: SCM for causal interventions, Hawkes processes for temporal dynamics, LLM agents for user responses
- Prospective simulation enables 'what-if' queries like shifting 30% budget or changing publish time by 6 hours
- SCM/LLM interface is an acknowledged weak spot—prompt interventions ≠ latent user interventions
Why It Matters
Lets marketers test budget and content changes in simulation before spending real dollars.