Food CPI Forecast with LSTM and Incidence Analysis

Overview

This project develops an integrated framework to forecast Ecuador’s Food Consumer Price Index (CPI-Food) and diagnose the subgroups exerting the greatest inflationary pressure. It combines deep learning (LSTM) for short-term time-series forecasting with nonlinear regression (Random Forest) for incidence analysis.

Author: Mateo Pilaquinga
Institution: Yachay Tech University, School of Mathematical and Computational Sciences

Objectives

1. Forecast: Predict CPI-Food values for a 3-6 month horizon using a Long Short-Term Memory (LSTM) network.
2. Diagnose: Identify key drivers of food inflation (e.g., bread, cereals, meat) using a Random Forest model.
3. Support Decision Making: Provide tools for public institutions and private agents to anticipate inflation and design stabilization policies.

Data Source

The dataset consists of official monthly series from the National Institute of Statistics and Census (INEC) of Ecuador.

• Period: January 2006 to October 2025 (238 observations).
• Target Variable: Food CPI (IPC-Alimentos).
• Features: Eleven COICOP food subindices (e.g., Oils & Fats, Meat, Fruits, Vegetables).

Methodology

1. Forecasting with LSTM

• Model: Multi-step LSTM network.
• Input: Sliding temporal windows of past observations (standardized).
• Output: Joint estimation of future values (3-6 months ahead).
• Performance: The LSTM achieves low error levels (RMSE = 1.18, MAPE = 1.01% on test set), capturing trends, seasonality, and shocks effectively.

2. Incidence Analysis with Random Forest

• Model: Random Forest regressor trained on monthly variation rates.
• Goal: Quantify the contribution of each food subgroup to the aggregate CPI variation.
• Findings: Bread and cereals, non-alcoholic beverages, and meat are the primary structural drivers, while fresh products (fruits, vegetables) show higher volatility but marginal aggregate impact.

Project Structure

├── data/
│   ├── processed/      # Processed data used for modeling (data.csv)
│   ├── ipc.csv         # Raw CPI data
│   └── productos.csv   # Raw product subindices data
├── src/                # Source code
│   ├── main_models.py          # Main entry point for the pipeline
│   ├── lstm_model.py           # LSTM model implementation
│   ├── random_forest_model.py  # Random Forest model implementation
│   ├── preprocessing.py        # Data cleaning and preparation modules
│   ├── evaluation.py           # Metrics and plotting utilities
│   └── compare_predictions.py  # Comparison script for validations
├── notebooks/          # Jupyter notebooks for exploration
│   ├── data_clean.ipynb        # Data cleaning process
│   ├── eda.ipynb               # Exploratory Data Analysis
│   └── future_eng.ipynb        # Feature engineering experiments
├── docs/               # Documentation
│   ├── clean_process.md        # Details on data cleaning
│   ├── description_variables.md # Variable dictionary
│   └── report_results.md       # Detailed model results
├── models/             # Directory for saved models (.h5, .pkl)
├── results/            # Generated metrics and predictions (.csv, .json)
├── figs/               # Generated plots and visualizations
└── requirements.txt    # Python dependencies

Installation

1. Clone the repository:

   git clone https://github.com/BrDenky/CFBPredic.git
   cd CFBPredic

2. Create a virtual environment (optional but recommended):

   python -m venv venv
   # Windows
   .\venv\Scripts\activate
   # Linux/Mac
   source venv/bin/activate

3. Install dependencies:

   pip install -r requirements.txt

Usage

To run the complete pipeline (preprocessing, training, evaluation, and visualization), execute the main script:

python src/main_models.py

This will:

1. Load and preprocess the data from data/processed/data.csv.
2. Train the LSTM model for forecasting.
3. Train the Random Forest model for feature importance.
4. Generate plots in figs/ and save metrics in results/.
5. Save trained models in models/.

Results Summary

Model	Metric	Value	Interpretation
LSTM	RMSE	1.18	High accuracy, error magnitude close to index scale (~1 point)
LSTM	MAPE	1.01%	Excellent relative accuracy (<5%)
Random Forest	RMSE	9.91	Used for explanatory purposes (incidence), not forecasting

Key Inflation Drivers (Random Forest Importance):

1. Bread and Cereals (~17%)
2. Non-alcoholic Beverages (~16%)
3. Meat (~14.6%)

License

MIT License (or specify appropriate license)