Time Series Forecasting

Project Overview

Developed as part of the Artificial Intelligence Techniques in Forecasting and Optimization in Business Systems (TIAPOSE) course at the University of Minho (MIEGSI), this project addresses a real-world decision-support problem in commercial retail management. Using daily visitor traffic data provided by an industry partner from a pilot deployment, the system combines time series predictive forecasting with prescriptive optimization to guide 7-day marketing campaign allocations and maximize business profitability.

The implementation strictly leveraged R and Professor Paulo Cortez’s rminer framework, integrating temporal forecasting techniques, scenario-based pre-processing, and web app deployment via R Shiny.

---

Technical Architecture & Methodology

The project strictly adhered to the CRISP-DM (Cross-Industry Standard Process for Data Mining) framework across six structured stages:

1. Business Understanding

• Defined the core objective: transform raw foot-traffic time-series data into actionable business intelligence by forecasting daily entrance counts and determining optimal marketing spend.
• Formulated the dual-stage framework: Stage 1 predicts daily visitor volumes, while Stage 2 feeds these predictions into an optimization algorithm to select the most profitable promotional campaign strategy for a target 7-day window.

2. Data Understanding & Exploratory Analysis (EDA)

• Conducted comprehensive exploratory analysis on historic visitor entry records provided by the IT pilot project.
• Identified temporal patterns, weekly seasonality, peak foot-traffic days, and variance trends across different time horizons.
• Examined data quality, cardinality, missing values, and statistical distribution properties to establish baseline expectations for regression performance.

3. Data Preparation & Scenario Engineering

• Feature Engineering: Formulated synthetic temporal attributes, combined collinear indicators, and engineered lag variables to capture autoregressive behavior.
• Outlier & Error Handling: Cleaned anomalies, imputing or adjusting extreme values that could distort regression boundaries.
• Scenario Testing: Designed three distinct pre-processing scenarios with varying feature subsets (selected through manual correlation matrix analysis and automated grid search experiments) to compare model sensitivity across pre-processed inputs.

4. Modeling & Backtesting Evaluation

• Model Families: Constructed univariate, multivariate, and hybrid time-series regression models using the rminer and forecast packages in R.
• Temporal Validation: Implemented rolling-window backtesting and time-aware cross-validation to prevent data leakage and ensure realistic out-of-sample performance evaluation on sequential time-series data.
• Performance Metrics: Tracked regression evaluation metrics (RMSE, MAE, MAPE) across each scenario to select the optimal model configuration.

5. Prescriptive Optimization Engine

• Fed the 7-day forecasted visitor numbers into a mathematical optimization script.
• Evaluated marketing campaign options against expected visitor traffic conversion rates to determine the optimal budget distribution that maximizes ROI while minimizing marketing acquisition cost.

6. Application Deployment

• Developed an interactive web application using R Shiny in RStudio, encapsulating model loading, dynamic prediction generation, and interactive optimization controls.
• Deployed the full production app on Shinyapps.io, enabling real-time scenario simulation and web-based decision making.

---

Key Challenges & Technical Insights

• Transitioning from Python to R: Migrated machine learning paradigms from Python/Scikit-Learn to R’s specialized data science library ecosystem (rminer, forecast).
• Time Series Backtesting: Mastered time-aware evaluation strategies (rolling window backtesting) to properly measure model generalization without exposing future temporal information.
• Predictive-to-Prescriptive Integration: Successfully linked a predictive regression engine directly into a downstream optimization function, converting raw numbers into strategic marketing decisions.
• End-to-End App Architecture: Built and published a functional web UI and RESTful processing logic in R Shiny from scratch.