---
title: "README"
author: "Gabriel Gonzalo Ojeda Cárcamo"
date: "7/1/2026"
output: pdf_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
```


## Overview

This project develops a **short-term demand forecasting solution** for industrial products operating under **high volatility and limited historical data** (40 monthly observations per product, no exogenous variables).

Classical time series models (ARIMA, ETS) are used as benchmarks and compared against a **feature-based LightGBM model** with shock-aware features.

---

## Business Problem

Forecasts support **monthly production planning and inventory decisions**.

- Over-forecasting → higher holding costs  
- Under-forecasting → stockouts and lost sales  

Because demand is irregular and shock-driven, the solution provides **probabilistic forecasts (P10–P50–P90)** to quantify uncertainty and support scenario-based planning.

---

##  Data & Constraints

- Monthly data (2021–2024)
- 40 observations per product
- No exogenous variables
- High volatility, non-Gaussian changes

These constraints require **robust and parsimonious models** with realistic validation.

---

##  Modeling Approach

**Benchmarks**
- ARIMA
- ETS (discarded due to poor performance under volatility)

**Final Model**
- LightGBM regression with engineered features:
  - Lagged demand (1, 2, 3, 6, 12)
  - Rolling averages (3, 6, 12)
  - Cyclical seasonality (sin/cos)
  - Shock features capturing abrupt changes

---

##  Validation

Models are evaluated using **walk-forward (expanding window) cross-validation**, retraining at each step and forecasting one month ahead.  
This setup mimics real deployment and avoids look-ahead bias.

---

##  Key Results (PR2 example)

| Model | MAPE |
|-----|------|
| ARIMA | 29% |
| ETS | 30% |
| LightGBM (basic) | 26% |
| **LightGBM + shock features** | **15%** |

→ **Forecast error reduced by 50%** relative to the ARIMA baseline.

---

##  Outputs

- Point forecasts
- Quantile forecasts: **P10 / P50 / P90**
- 6-month horizon forecasts for each product

---

##  Notes

- Small sample size (40 obs)
- No exogenous drivers
- Short-term forecasting only (1–6 months)

Deep learning models were intentionally excluded due to overfitting risk under these constraints.

---

## Project Structure

src/

  - features.R
  - walk_forward.R
  - models_lgbm.R
  - forecast_quantiles.R
  
report/

  - time_series_forecasting.pdf
  
README.md

## Contact

**Gabriel Gonzalo Ojeda Cárcamo**  
Economist | MSc Statistics (candidate)  
Data Analytics