Project Overview

This project investigates time series forecasting for paediatric emergency department attendances. The dataset contains daily ED attendance counts from 01/04/2014 to 19/02/2017. The objective is to develop and compare forecasting models to predict attendances for the next 28 days and evaluate their performance.

Goals

• Understand seasonal trends in paediatric ED attendances (weekly & monthly patterns).
• Develop and evaluate forecasting models (Naive, ARIMA, and Prophet).
• Compare models using performance metrics (MAE and Winkler scores for prediction intervals).
• Make recommendations for staff planning based on forecast accuracy.

---

Key Files

Jupyter Notebooks

• 00_introduction.ipynb – Project overview.
• 01_data_exploration.ipynb – Exploratory data analysis (EDA), visualising trends, seasonality and outliers.
• 02_naive_model.ipynb – Implements Naive 1 and Seasonal Naive models as benchmarks.
• 03_arima_model.ipynb – Develops and evaluates an ARIMA models for forecasting.
• 04_prophet_model.ipynb – Implements Facebook Prophet models for capturing trends and seasonality.
• 05_final_report.ipynb – Main report, insights and recommendations.
• 06_references.ipynb – References.

Folder Structure

project/
├── README.md
├── binder/
│   └── environment.yml
├── 00_introduction.ipynb
├── 01_data_exploration.ipynb
├── 02_naive_model.ipynb
├── 03_arima_model.ipynb
├── 04_prophet.ipynb
├── 05_final_report.ipynb
├── 06_references.ipynb
├── images/
│   └── auto_arima_graph.png
│   ├── naive_graph.png
│   ├── box_plots.png
│   ├── prophet_predictions.png
│   └── prophet_graph.png
├── input/
│   └── model_runs.keras
├── paediatrics_train.csv
└──  ts_utils.py

Data Files

• paediatrics_train.csv – The dataset containing daily paediatric ED attendances.

Scripts

• ts_utils.py – Utility functions for time series preprocessing, modeling and evaluation.

---

Methodology

1. Data Exploration

• Visualising monthly and weekly seasonality.
• Identifying trends and outliers.
• Applying stationarity tests (ADF, KPSS).

2. Forecasting Models

Model	Description
Naive	Baseline models for benchmarking, investigated Naive1 & SeasonalNaive
ARIMA	Classical time series model (Auto-Regressive Integrated Moving Average)
FB Prophet	Forecasting model designed for handling seasonality and holidays

3. Model Evaluation

• MAE (Mean Absolute Error) – Measures average forecast error.
• Winkler Score (80% & 90%) – Evaluates prediction intervals.

---

Installation & Setup

To run the notebooks, set up the required environment:

1. Conda Environment

conda env create -f binder/environment.yml

conda activate hds_forecast

Author: Kayleigh Haydock