Detailed Documentation

“FutureOrders - Forecasting Demand with AI”

Name: Kartikey Chaurasia & Syed Adnan Rashid

Course: MBA (AI & DS)
1

FutureOrders - Forecasting Demand with AI

Introduction

Demand forecasting is one of the main activities of supply chain management that helps
companies predict customer demand, manage inventory, and make optimal allocation of
resources. With greater global market complexity, accurate demand forecasting has become
more critical than ever. Companies employ forecasting to prevent stockouts, reduce excess
inventory, and drive overall operating efficiency to its maximum level. A well-functioning
demand forecasting system generates customer satisfaction through ensured product
availability while reducing holding costs concurrently.
Traditional demand forecasting techniques, such as moving averages and exponential
smoothing, have been used for decades. However, these approaches often fail to capture
complex patterns, especially in datasets exhibiting high seasonality, trends, or sudden
fluctuations. In response to these limitations, modern machine learning and deep learning
techniques have emerged as powerful alternatives. By leveraging vast amounts of historical
data, these advanced methods can identify intricate relationships and patterns that traditional
models might overlook.
Machine learning-based demand forecasting integrates a variety of algorithms, including
decision trees, support vector machines, neural networks, and ensemble methods like
XGBoost. Each of these techniques has its strengths, depending on the nature of the data and
the specific business context. Additionally, hybrid models combining statistical and machine
learning approaches have gained popularity due to their ability to balance interpretability and
predictive power.
In this project, we focus on forecasting daily demand using the "Daily Demand Forecasting
Orders" dataset. We explore a range of forecasting models, from statistical methods like
ARIMA to machine learning techniques such as LSTM and XGBoost. By evaluating the
performance of these models, we aim to identify the most effective approach for predicting
demand fluctuations accurately. This study offers important information on the application of
machine learning in demand forecasting and different approaches with advantages and
disadvantages.
Machine learning demand forecasting involves a range of algorithms, such as decision trees,
support vector machines, neural networks, and ensemble techniques such as XGBoost. Each
has its advantages, depending on the type of data and the particular business situation.
Furthermore, hybrid models where statistical and machine learning methods are used together
have also become widely applied because they can find the optimal trade-off between
interpretability and predictive capability.
2

FutureOrders - Forecasting Demand with AI

Literature Review
It has been one of the widely researched topics in both academia and industry because of its
importance in supply chain management. There are multiple models proposed for better
predictive accuracy, but they start from classic statistical approaches to more recent machine
learning techniques. These include the following methods:

Traditional Time Series Models:

• Autoregressive Integrated Moving Average (ARIMA): One of the most widely used
time series forecasting methods, ARIMA effectively captures linear dependencies and
seasonal patterns. However, it assumes stationarity and struggles with highly volatile data.
• Exponential Smoothing Methods: These models, e.g., Holt-Winters, use exponentially
decreasing weights on previous observations and are therefore good for following trends
and seasonality in short-term forecasts.
• Moving Average (MA) and Seasonal Decomposition of Time Series (STL): Both these
methods are effective for trend and seasonality detection in general but weak in predicting
in long-term forecasting.

Machine Learning-Based Approaches:

• Random Forests & Gradient Boosting: Ensemble techniques involving multiple
decision trees to improve forecasting accuracy and prevent overfitting. The popular
gradient boosting algorithm XGBoost has found extensive use in demand forecasting due
to its stability and efficiency.
• Support Vector Machines (SVMs): Applicable in high-dimensional prediction problems,
SVMs offer insensitivity to outliers and are able to handle data non-linearity.
3

FutureOrders - Forecasting Demand with AI

• Artificial Neural Networks (ANNs): Through imitation of human thought processes,
ANNs are capable of learning intricate relationships in pattern demands. But they need
considerable training data and hyperparameter tuning.

Deep Learning Models:

• Long Short-Term Memory (LSTM) Networks: A dedicated variant of recurrent neural
networks (RNNs), LSTMs are designed especially to manage long-term dependencies in
sequential data, e.g., demand forecasting.
• Convolutional Neural Networks (CNNs): Although originally used for image
processing, CNNs have been used for time series forecasting too, and efficiently identified
local dependencies in demand fluctuations.
• Hybrid Models (ARIMA + LSTM): Blending statistical models and deep learning
models has been effective in determining both linear and non-linear patterns in demand
forecasting, leading to better predictive precision.

Challenges in Demand Forecasting:

• Data Volatility and Uncertainty: Demand patterns are subject to several external
influences, including market trends, economic conditions, and seasonal fluctuations, and
hence forecasting is extremely uncertain.
• Overfitting and Data Limitations: Machine learning models tend to overfit noisy data,
resulting in poor generalization. Good train-test splitting and cross-validation are critical
to avoid this problem.
• Interpretability vs. Accuracy Trade-off: Deep learning models, while more accurate, are
less interpretable compared to traditional statistical methods, thus making it more
challenging to make decisions.

Emerging Trends in Demand Forecasting:

• Incorporation of External Data Sources: Including live market trends, customer
opinion, and macroeconomic signals can help make demand forecasts more accurate.
• Reinforcement Learning Approaches: Adaptive forecasting methods using
reinforcement learning are gaining traction, allowing models to adjust dynamically to
changing demand patterns.
• Quantum Computing Applications: Quantum computing's ability to optimize demand
forecasting models is being investigated, with potential advances in computational
efficiency and predictability.
4

FutureOrders - Forecasting Demand with AI

In conclusion, the development of demand forecasting models has evolved from the use of
classical statistical methods to sophisticated machine learning and deep learning models. Each
method, however, has its advantages and limitations, while hybrid models and external data
incorporation are promising fields of study for the future.

FutureOrders - Forecasting Demand with AI

Methodology

3.1 Data Collection The dataset consists of daily order records, including various features
influencing demand. Key attributes include:
• Week of the month (First, second, third, fourth, or fifth week)
• Day of the week (Monday to Friday)
• Non-urgent order count
• Urgent order count
• Order type (A, B, and C)
• Fiscal sector orders
• Orders from the traffic controller sector
• Banking orders (Three categories)
• Target variable: Total orders

3.2 Data Preprocessing

• Handling missing values and outliers
• Converting categorical and time-based features into numerical representations
• Normalization and scaling for machine learning models
6

FutureOrders - Forecasting Demand with AI

• Splitting data into training and testing sets

3.3 Model Selection & Training Multiple regression-based and tree-based machine learning
models were developed and validated for demand forecasting. The models are as below:
• Linear Regression: A basic model that identifies linear trends with respect to demand.
• Ridge Regression & Lasso Regression: Regularized linear models reducing overfitting.
• Decision Tree Regressor: A non-linear model that separates data based on feature values.
• Random Forest Regressor: An ensemble model consisting of multiple decision trees that
enhance the accuracy of prediction.
• Gradient Boosting & XGBoost Regressors: Advanced boosting models known for their
strong predictive power and ability to handle complex relationships..

3.4 Cross-Validation & Hyperparameter Tuning To prevent overfitting in the model, we

performed k-fold cross-validation. Thus, we divided our dataset into several training and
validation sets. We have tuned our hyperparameters by performing Grid Search and
Randomized Search for fine-tuning model performance.

3.5 Feature Importance Analysis To understand which factors influence demand the most,
we analyzed feature importance scores from tree-based models (Random Forest, Gradient
Boosting, and XGBoost). These insights can help businesses prioritize key demand drivers in
decision-making.

3.7 Model Evaluation Performance measures utilized to compare the models are:
• Mean Absolute Error (MAE): Measures the average magnitude of errors in predictions.
• Mean Squared Error (MSE): Computes the average squared difference between actual
and predicted values.
• Root Mean Squared Error (RMSE): Provides an interpretable measure of prediction
accuracy.
7

FutureOrders - Forecasting Demand with AI

Results and Key Findings

Model Performance Comparison:

• Linear models (Linear, Ridge, and Lasso Regression) gave a fair baseline but not for
predicting non-linear trends.
• Decisions Trees were good but overfitting without tuning.
• Random Forest and Gradient Boosting techniques showed better accuracy compared to
individual models.
• XGBoost performed better than other models because of its effective management of
feature interactions and regularization methods.

Performance Metrics:
• Best Model: XGBoost
o Mean Absolute Error (MAE): 24.25
o Root Mean Squared Error (RMSE): 40.67
o R² Score: 0.84
o Cross-Validation R² Mean: 0.9041

Best Hyperparameters for XGBoost:

• Subsample: 0.7
• Number of Estimators: 300
• Max Depth: 3
• Learning Rate: 0.05
• Column Sample by Tree: 0.7

Feature Importance Analysis:

• Non-urgent orders had the highest impact (42.53%).
• Order Type B (25.89%) and Order Type C (13.50%) were also significant predictors.
• Urgent orders contributed 11.51% to demand forecasting.
• Banking and fiscal sector orders had minor influence compared to other variables.
8

FutureOrders - Forecasting Demand with AI

Key Observations:
• Seasonality significantly influenced demand variations.
• Holidays and weekends showed higher demand fluctuations.
• Feature importance analysis revealed that past demand trends and product types were
strong predictors of future demand.

Overall Interpretation and Conclusions:

This research proved that machine learning algorithms, XGBoost in this instance, is a strong
tool for demand forecasting. XGBoost had an R² of 0.84 and cross-validation R² of 0.9041,
hence making accurate predictions while highlighting the major drivers of demand. The drivers
with the most influence in forecasting demand were the types of orders, urgent orders, and non-
urgent orders. This study identifies the need for data-driven methodologies in supply chain
management, which enables companies to maximize inventory planning and resource
deployment. Deep learning methods and other external factors may be further studied in future
studies to increase the accuracy of forecasts.
9

FutureOrders - Forecasting Demand with AI

Conclusion
This study demonstrates the successful application of machine learning techniques for demand
forecasting, emphasizing the superiority of the XGBoost model in predicting daily orders with
high accuracy. By leveraging advanced time series analysis, feature engineering, and rigorous
model evaluation, we identified key demand drivers, such as non-urgent orders, order types,
and urgent orders, which significantly impact total demand. The results indicate that traditional
linear models struggle with complex demand patterns, while tree-based models, particularly
XGBoost, offer enhanced predictive power due to their ability to capture intricate feature
interactions. Our findings provide valuable insights for businesses seeking to optimize
inventory management and resource allocation, helping them reduce stock shortages and
excess inventory while improving overall operational efficiency. Additionally, the study
underscores the importance of cross-validation and hyperparameter tuning in ensuring model
robustness. Although our model demonstrates strong predictive capability, external factors such
as market trends, economic conditions, and policy changes could still impact demand,
highlighting the need for continuous model monitoring and adaptation. Future research could
explore deep learning approaches, hybrid models, or the integration of real-time data streams
to further refine forecasting accuracy. In all, the project reaffirms the importance of data-led
forecasting in contemporary supply chain management and business decision-making,
affirming how machine learning has the potential to transform demand planning strategies.

10

FutureOrders - Forecasting Demand with AI