Multi-Agent Causal Reasoning Framework: Optimizing Advertising Incrementality via Daily Budget Allocation Under a Fixed Lifetime Budget

Paul Lo, Jing Pan, Jingxi Yu, Evelyn Wu, Erte Pan, Simon Hollands

Keywords: Budget allocation, Reach response modeling, Incrementality, Multi-agent planning, Causal Inference

TL;DR: We propose a Multi-Agent Causal Reasoning Framework using incremental reach as a causal surrogate for budget optimization. By modeling reach curves and diminishing returns, it reallocates spend to maximize incrementality under a fixed budget.

Abstract:

We study a practical approach for improving advertising incrementality by reallocating spend across days within a fixed lifetime budget $B$ throughout the campaign flight. Our key observation is that the marginal gains in incremental unique reach—additional distinct users reached due to advertising interventions—vary systematically over time (e.g., weekly seasonality) and diminish as spend increases. Consequently, a non-uniform daily budget split can increase total incremental unique reach, improving both incrementality outcomes and the statistical power of measurement.We propose the Multi-Agent Causal Reasoning Framework, which plans a multi-day budget schedule under a fixed lifetime constraint using learned reach response models trained on historical campaigns. The system comprises four specialized agents:
* $\text{Lifetime Budget Splitter (LBS): The Planning Agent}$ that generates candidate daily budget schedules $\{b_i\}_{i=1}^N$ to explore the decision space.
* $\text{Daily Reach Estimator (DRE): The Estimation Agent}$ that models the causal environment to predict daily incremental reach $r_i$ given campaign context and candidate budgets.
* $\text{Overall Reach Aggregator (ORA): The Governing Agent}$ that resolves resource competition by synthesizing per-day reach predictions to rank global schedules.
* $\text{Outcome Simulation Module (OSM): The Simulation Agent}$ that performs counterfactual simulations to estimate end-to-end incrementality performance under explicit reach-to-outcome mapping assumptions.

In an offline evaluation, the reach model achieves $R^2 = 0.78$ with a Mean Absolute Percentage Error (MAPE) of $18.8\%$. An end-to-end case study suggests a potential reduction in Cost Per Incremental Conversion (CPiC) of approximately $63\%$ under the proposed reach-to-CPiC mapping.

Paper Type: Full (minimum of 10 pages and a maximum of 16 excluding references)

Poster Opt In: Yes, I'm open to having my submission accepted as a poster (leave blank if you are submitting a poster, or if you DON'T want your submission to be accepted as a poster instead of a full or short paper)

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 8