🤖 machine learning

Hybrid Deep Learning Approach for Coupled Demand Forecasting and Supply Chain Optimization

This paper proposes HAF-DS, a hybrid AI framework that integrates LSTM-based demand forecasting with mixed integer linear programming optimization to significantly improve prediction accuracy and reduce operational costs in volatile textile and PPE supply chains.

Original authors: Nusrat Yasmin Nadia, Md Habibul Arif, Habibor Rahman Rabby, Md Iftekhar Monzur Tanvir, Md. Jakir Hossen, M. F. Mridha

Published 2026-04-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nusrat Yasmin Nadia, Md Habibul Arif, Habibor Rahman Rabby, Md Iftekhar Monzur Tanvir, Md. Jakir Hossen, M. F. Mridha

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are running a massive, bustling clothing factory that also makes emergency medical gear (like masks and gowns). You have two big problems:

The Crystal Ball Problem: You need to guess how many shirts or masks people will buy next month. If you guess wrong, you either have too much stuff gathering dust (wasting money) or too little (angry customers).
The Puzzle Problem: Once you have a guess, you need to figure out the best way to get the materials, hire the workers, and ship the goods without spending a fortune.

The Old Way (The "Silo" Approach):
In the past, companies treated these as two separate jobs.

Team A (The Forecasters): They used math to guess future sales. They were great at guessing numbers, but they didn't care about the cost of shipping or if the factory was actually running out of fabric.
Team B (The Optimizers): They took Team A's guess and tried to build a plan. But because Team A's guess was just a number, Team B often had to make up rules like "assume shipping is always fast" or "assume prices never change."

The Result: Team A guessed "We will sell 1,000 units." Team B made a plan based on that. But in reality, shipping was slow, and the plan failed. The two teams were like two people trying to drive a car while looking in opposite directions.

The New Solution: The "Super-Brain" (HAF-DS)

This paper introduces a new system called HAF-DS. Think of it not as two separate teams, but as a single, super-intelligent brain that does both jobs at the same time.

Here is how it works, using a simple analogy:

1. The "Smart Watch" (The LSTM Part)

Imagine your smartwatch doesn't just count your steps; it knows your habits, the weather, and your schedule.

In this system, the LSTM (a type of AI) is like that smartwatch. It looks at years of sales data, holidays, and trends.
But here's the magic: It doesn't just guess the number. It learns to guess the number specifically in a way that helps the factory save money. It learns, "Oh, if I predict a huge spike in demand, the shipping costs will go up, so I should be a little more careful with my prediction."

2. The "Chess Master" (The Optimization Part)

Now, imagine a Chess Master who doesn't just move pieces randomly but calculates the best move to win the game.

This part of the system looks at the "Smart Watch's" prediction and immediately asks: "Okay, if we need 1,000 units, what is the cheapest way to get them? Which supplier is reliable? How much should we keep in the warehouse?"
It considers real-world limits: "We can't order more than the truck can carry," or "This supplier is slow."

3. The "Handshake" (The Hybrid Link)

This is the most important part. In old systems, the Smart Watch and the Chess Master never talked to each other after the game started.

In this new system, they are holding hands.
If the Chess Master realizes the plan is too expensive, it sends a signal back to the Smart Watch: "Hey, your prediction led to an expensive plan. Try to adjust your guess next time to help us save money."
If the Smart Watch sees a pattern that usually causes a stockout, it tells the Chess Master: "Get ready, we might need more stock."

They learn together. The system gets better at guessing and better at planning simultaneously.

What Happened When They Tried It?

The researchers tested this "Super-Brain" on real data from textile factories and PPE manufacturers. Here is what they found, translated into everyday terms:

Fewer Mistakes: The system made fewer bad guesses about how much people would buy. It was about 15% more accurate than the old methods.
Less Waste: Because the guesses were better, the factories didn't have to store as much extra "just in case" inventory. They saved about 5% on storage costs.
No Empty Shelves: The biggest win was avoiding "stockouts" (running out of product). The system reduced the times they ran out of stock by 27.5%. That means customers got their orders on time much more often.
Happier Customers: The "Service Level" (how often they could say "Yes, we have it!") jumped from 95.5% to 97.8%.

The Bottom Line

Think of this paper as the difference between a navigator who just points to a destination and a driver who knows how to steer the car.

Old Way: The navigator says, "Drive North." The driver drives North but hits a pothole and breaks the car.
New Way (HAF-DS): The navigator and driver are the same person. They see the pothole coming, adjust the route while predicting the destination, and arrive safely, quickly, and with less fuel.

This new AI framework proves that when you combine predicting the future with planning for the present, you don't just get a better guess—you get a better, cheaper, and more reliable business.

1. Problem Statement

The paper addresses a critical disconnect in modern supply chain management, particularly in volatile industries like textiles and Personal Protective Equipment (PPE) manufacturing.

The Gap: Traditional approaches treat demand forecasting (predictive analytics) and supply chain optimization (prescriptive analytics) as sequential, loosely coupled processes. Forecasting models are trained solely to minimize prediction error (e.g., MAE, RMSE) without considering downstream operational costs. Conversely, optimization models often rely on static forecasts and simplifying assumptions (e.g., deterministic lead times) that fail to capture real-world variability.
The Consequence: This separation leads to suboptimal decisions, resulting in overstocking, stockouts, inflated inventory costs, and reduced service levels, especially during crises like pandemics or logistics disruptions.
The Goal: To develop a unified framework that jointly optimizes forecasting accuracy and operational efficiency, ensuring that predictive insights directly drive cost-effective and service-oriented supply chain policies.

2. Methodology: The HAF-DS Framework

The authors propose the Hybrid AI Framework for Demand–Supply Forecasting and Optimization (HAF-DS). Unlike sequential "predict-then-optimize" models, HAF-DS employs an end-to-end differentiable architecture where the forecasting and optimization modules are trained simultaneously.

A. Core Architecture

The framework consists of two tightly coupled modules:

Demand Forecasting Module:
- Architecture: A stacked Long Short-Term Memory (LSTM) network.
- Input: Time-series sales data (sliding window of 30 days) and categorical features (SKU, Supplier, Seasonality) processed via learned embedding layers.
- Function: Captures temporal dependencies and contextual demand patterns to predict future demand ( $\hat{y}_{t+1}$ ).
Supply Chain Optimization Module:
- Architecture: A fully connected neural network (FCNN) acting as a differentiable optimization layer.
- Input: Predicted demand ( $\hat{y}_{t+1}$ ) and supply-side constraints (stock levels, lead times, supplier reliability, defect rates).
- Function: Generates optimized actions ( $o_{t+1}$ ), such as order quantities and supplier selection, to minimize total operational cost while satisfying constraints.

B. Key Technical Innovations

Gradient Coupling (Differentiable Optimization): The core novelty is the joint loss function that allows gradients from the supply-side optimization loss to flow backward into the forecasting network.
- Total Loss: $L_{total} = \lambda_1 L_{forecast} + \lambda_2 L_{supply}$
- Mechanism: The forecasting parameters ( $\theta$ ) are updated not just to minimize prediction error, but to minimize the downstream operational cost ( $L_{supply}$ ). This creates a feedback loop where the model learns to predict demand in a way that is "operationally feasible" and cost-efficient.
Embedding-Based Feature Representation: Categorical variables (e.g., Supplier ID, Product Type) are mapped to dense vector spaces via trainable embeddings, allowing the model to learn latent relationships between suppliers, SKUs, and time patterns that one-hot encoding misses.
Stochastic Regularization: To handle uncertainty, predicted demand is sampled from a Gaussian distribution ( $\hat{y}_{t+1} \sim N(\mu, \sigma^2)$ ), enabling the optimization module to adapt to confidence-weighted forecasts.

C. Training Strategy

Data: Utilized two public datasets: a Textile Sales Dataset (demand) and a Supply Chain Dataset (operations).
Preprocessing: Included mean/mode imputation, Min-Max normalization, and data augmentation (Gaussian noise injection) to improve robustness.
Optimization: Trained using the Adam optimizer with early stopping. The balancing weights ( $\lambda_1, \lambda_2$ ) were tuned via grid search to ensure neither module dominated the learning process.

3. Key Contributions

Unified Predictive-Prescriptive Architecture: Introduces a fully integrated framework where forecasting and optimization are jointly trained, bridging the gap between predictive accuracy and operational decision-making.
End-to-End Gradient Flow: Demonstrates how supply-side constraints and cost objectives can directly influence the learning of demand forecasting parameters, a significant departure from traditional two-stage approaches.
Contextual Embeddings: Proves that learned embeddings for categorical supply chain variables significantly enhance both forecasting precision and downstream optimization performance.
Comprehensive Evaluation: Provides extensive benchmarking against classical statistical models (ARIMA, Prophet), deep learning baselines (LSTM, GRU, Transformers), and reinforcement learning agents across multiple metrics.

4. Experimental Results

The framework was evaluated on textile and PPE datasets, comparing the proposed HAF-DS against baselines like ARIMA, LSTM, GRU, and Temporal Fusion Transformers (TFT).

A. Forecasting Accuracy

On the combined dataset, HAF-DS achieved significant reductions in error metrics compared to the strongest deep learning baseline (GRU):

MAE: Reduced from 15.04 to 12.83 (14.7% improvement).
RMSE: Reduced from 19.53 to 17.11 (12.4% improvement).
MAPE: Reduced from 9.5% to 8.1% (14.7% improvement).
Note: The model outperformed even advanced Transformer-based models (TFT, Informer).

B. Supply Chain Optimization Performance

Coupling forecasting with optimization led to tangible operational gains:

Inventory Cost: Decreased by 5.4%.
Stockout Rate: Reduced by 27.5% (from 4.0% to 2.9%).
Service Level: Increased from 95.5% to 97.8%.
Total Operational Cost: Reduced by 4.6%.

C. Ablation Studies

Joint Training: Removing the joint training (training modules separately) resulted in a ~19.5% increase in MAE and a ~19.3% increase in inventory costs, confirming the necessity of the unified approach.
Embeddings: Removing embeddings increased MAE by ~9.7%, highlighting their importance for contextual learning.
Sensitivity: The model remained robust to variations in learning rates and batch sizes, though extreme imbalances in loss weighting ( $\lambda$ ) degraded performance.

5. Significance and Impact

Theoretical Advancement: The paper formalizes the transition from "predict-then-optimize" to "predict-and-optimize" via differentiable coupling, offering a mathematically rigorous approach to aligning learning objectives with operational costs.
Practical Applicability: The framework is designed to be domain-agnostic and scalable. It can be integrated into existing ERP or SCM systems via modular APIs.
Resilience: By explicitly modeling supply-side constraints and uncertainty, the system is better equipped to handle volatile markets (e.g., pandemics, logistics shocks), reducing the risk of stockouts and overstocking.
Interpretability: The inclusion of attention mechanisms and constraint sensitivity analysis allows managers to trace decisions back to specific drivers (e.g., supplier reliability or seasonal trends), fostering trust in AI-driven recommendations.

In conclusion, the HAF-DS framework demonstrates that integrating deep learning forecasting with mathematical optimization in a single, differentiable pipeline yields superior accuracy and operational efficiency compared to traditional sequential methods, providing a scalable solution for modern, volatile supply chains.