[Paper Review] Multi-stream RNN for Merchant Transaction Prediction

Zhongfang Zhuang, Chin‐Chia Michael Yeh+3 more|arXiv (Cornell University)|Jul 25, 2020

Time Series Analysis and Forecasting9 references4 citations

TL;DR

This paper proposes a multi-stream RNN model that predicts multivariate, multi-step merchant transaction patterns by processing daily time series data through separate RNN streams, then fusing them via merge layers and a weekly RNN for long-term forecasting. The model outperforms state-of-the-art methods on real-world transaction data, demonstrating superior accuracy in predicting future transaction trends across multiple merchant categories.

ABSTRACT

Recently, digital payment systems have significantly changed people's lifestyles. New challenges have surfaced in monitoring and guaranteeing the integrity of payment processing systems. One important task is to predict the future transaction statistics of each merchant. These predictions can thus be used to steer other tasks, ranging from fraud detection to recommendation. This problem is challenging as we need to predict not only multivariate time series but also multi-steps into the future. In this work, we propose a multi-stream RNN model for multi-step merchant transaction predictions tailored to these requirements. The proposed multi-stream RNN summarizes transaction data in different granularity and makes predictions for multiple steps in the future. Our extensive experimental results have demonstrated that the proposed model is capable of outperforming existing state-of-the-art methods.

Motivation & Objective

• To address the challenge of predicting multivariate, multi-step transaction patterns for merchants in real-time payment systems.
• To model interdependencies among transaction features (e.g., volume, amount, average) across multiple time steps.
• To improve long-term forecasting accuracy by capturing daily and weekly temporal patterns using a multi-stream architecture.
• To provide a scalable and flexible deep learning framework for transaction monitoring and anomaly detection in payment ecosystems.
• To validate the model’s effectiveness on real-world aggregated transaction data from diverse merchant categories.

Proposed method

• The model uses separate RNNs (LSTM or GRU) for each day of the week to process hourly transaction data, capturing daily patterns.
• Hourly outputs from each daily RNN are aggregated at the same hour across days using merge layers to form a unified temporal representation.
• A weekly RNN processes the 24 merged time steps (one per hour) to learn high-level weekly patterns and dependencies.
• Fully connected layers generate predictions for each of the next 24 hours, with individual heads for each multivariate feature.
• A shrink stacked RNN variant is introduced to improve training stability and convergence, especially for deeper architectures.
• Dropout with 0.2 rate is applied after each RNN to prevent overfitting.

Experimental results

Research questions

• RQ1 Can a multi-stream RNN architecture effectively model multivariate, multi-step transaction predictions across different merchant types?
• RQ2 How does modeling daily patterns through separate RNN streams improve long-term forecasting compared to single-stream RNNs?
• RQ3 To what extent does the inclusion of a shrink stacked RNN improve model performance and convergence?
• RQ4 How well does the model generalize across diverse merchant categories such as restaurants, medical services, and retail?
• RQ5 Does the model preserve temporal dependencies across multiple future time steps better than existing RNN-based baselines?

Key findings

• The multi-stream RNN with shrink stacked RNN outperformed all baseline methods in 19 out of 24 experimental setups based on non-normalized RMSE.
• The model achieved better normalized RMSE in 17 out of 24 setups when using shrink stacked RNN, indicating improved generalization.
• For restaurant merchants, the model’s predictions closely matched ground truth in density plots, with overlapping distributions at the 23rd future hour.
• Visual analysis showed that the MS-RNN predicted more accurate patterns than the SOTA RNN, especially for long-term forecasts (e.g., hour 23).
• Increasing network depth improved shape modeling (higher normalized RMSE) but slightly degraded RMSE for LSTM and significantly for GRU, indicating trade-offs in architecture design.
• The model demonstrated robust performance across diverse merchant categories, including restaurants, medical services, and retail, with consistent improvements over baselines.