Executive Guide to Forecasting in Manufacturing, Part 5: Forecasting Methods

Every manufacturer eventually asks the same question:

Which forecasting method should we use?

Should we trust the simple average, adopt statistical models, or lean into AI? The search for the “right” technique often overshadows a more practical truth: the best method depends on the business context.

In Demand Forecasting for Executives and Professionals by Stephan Kolassa, Bahman Rostami-Tabar, and Enno Siemsen dedicate several chapters to this question. They trace the evolution of forecasting methods from straightforward rules of thumb to formal statistical models like Exponential Smoothing and ARIMA.

What emerges is not a contest between methods, but a continuum of capability. Each approach serves a purpose: simple methods provide clarity, statistical models add rigor, and AI brings adaptability. The real skill lies in knowing when each belongs in your planning process.

Key Insights from the Book

• Simple methods remain powerful baselines for accuracy and trust.
  
• Exponential Smoothing combines adaptability with stability and performs with high accuracy.
  
• ARIMA brings structure and analytical depth to time-series data.
  
• Forecasting progress comes from layering methods, not replacing them.
• Forecasting maturity lies in progression, not replacement. Each method adds a layer of sophistication while retaining the simplicity that grounds good judgment.

Simple Forecasts: Straightforward, Often Efficient

The book begins with the simplest forecasting ideas: the historical mean, the naïve forecast, and the seasonal naïve forecast.
Each method uses recent data without parameters or training: averaging past demand, repeating the last observation, or repeating the value from the same period in the previous cycle.

The authors argue these should always be the starting point. As the authors note, roughly half of all forecasts fail to beat a simple no-change model (Morlidge, 2014b).

For manufacturers, simple methods offer reference forecasts for benchmarking more advanced models and setting expectations about achievable accuracy.

Exponential Smoothing:  Adaptivity Without Complexity

Exponential Smoothing (ES) is introduced as one of the most reliable and scalable forecasting techniques. It updates forecasts by weighting recent observations more heavily while still accounting for older data.

Key strengths:

• Captures gradual changes without overreacting to noise.
  
• Scales efficiently across large product portfolios.
  
• Remains interpretable for planners and executives.
  

For manufacturing planners, it offers a dependable middle ground: robust enough for large product portfolios, yet still interpretable for operational review meetings.

ARIMA Models — Structure and Statistical Discipline

ARIMA (Autoregressive Integrated Moving Average) are another core family of statistical forecasting models. ARIMA models describe demand as a function of its past values and past errors, adjusted through differencing to remove trends.

They are best suited for:

• Series with short-term correlations and limited seasonality.
  
• Situations where understanding underlying demand structure matters more than automated scale.

In large forecasting comparisons, Exponential Smoothing generally performs better, though ARIMA remains useful for diagnosing how patterns evolve over time.

Artificial Intelligence and Machine Learning: Power and Complexity

In later chapters, the authors shift from classical statistical approaches to Artificial Intelligence (AI) and Machine Learning (ML). A diverse set of techniques that have transformed how many organizations approach forecasting.

These include:

• Neural networks and deep learning, where models learn patterns by mimicking neural processes in the brain.
  
• Recurrent and LSTM networks, which handle sequential data and capture “memory” over time.
  
• Tree-based methods and Random Forests, which split data into decision rules and combine hundreds of small models.
  
• Boosting methods such as LightGBM and XGBoost, which iteratively improve predictions by modeling residual errors.
  

AI and ML methods excel at detecting nonlinear relationships and complex interactions that traditional models can’t. They can also integrate a wide range of external predictors—weather, pricing, promotions, macroeconomic signals—making them well suited for rich, high-volume data environments.

However, the authors are careful to temper expectations. These models are data- and resource-intensive. They require large datasets, extensive computation, and careful governance to prevent overfitting.

In summary, the book’s stance is balanced: AI and ML can outperform traditional methods, especially where demand drivers are nonlinear or where vast data volumes are available. But their value depends on fit. When applied without clear data needs or operational alignment, they can become expensive answers to simple problems.

For mature organizations, the lesson is not to reject or romanticize AI, but to apply it where the data and business case justify the investment, ensuring that models remain explainable, scalable, and ultimately actionable.

Quantics Perspective

In enterprise manufacturing, the question of “which forecasting method to use” is rarely straightforward. Demand patterns are difficult to predict: customers order in batches, portfolios span multiple business units and channels, and observed data is often influenced by internal processes such as order promising or capacity constraints.

As a result, what looks like demand is often a mix of market signals and operational decisions — adding complexity to forecasting.

No Single Method Fits All

In this environment, robustness and accuracy are critical. No single method consistently outperforms others across all use cases. Each has its strengths:

• Simple methods provide transparency and strong baselines
• Statistical models offer stability and scalability
• AI / ML methods bring flexibility but require more data and governance

At Quantics, we use an ensemble approach, combining forecasts from 60+ methods across statistical, machine learning, and deep learning models.

• Leverages complementary strengths
• Reduces reliance on a single model
• Improves robustness across changing demand patterns

Leveraging Structure Across Levels

Rather than forecasting only at the most granular level, Quantics applies hierarchical forecasting, generating predictions across multiple levels and reconciling them.

This approach:

• Uses stable patterns at aggregated levels
• Retains local signals at SKU level
• Improves accuracy while reducing bias

Adapting to Change

Manufacturing environments are dynamic. Market conditions, product lifecycles, and internal processes evolve.

AI-driven approaches help to:

• Incorporate recent data
• Integrate additional signals
• Adjust dynamically to structural changes

Backtesting: Measuring What Works

Robust model selection requires systematic backtesting — not just a single historical fit.

Evaluate over time

• Use multiple rolling forecasts to assess stability across periods

Assess across levels

• Measure accuracy and bias beyond SKU level to support operational decisions

Test across horizons

• Short-term (up to 3 months): production and inventory
• Longer-term (up to 18 months): tactical planning
  ‍

At Quantics, forecasting is not about selecting a single “best” method, but about building a robust system. Combining methods, learning across levels, and continuously adapting to change. This enables more resilient, data-driven planning.

Practical Takeaways

• Start simple and benchmark everything.
  Use naïve and simple models as a baseline — many advanced methods fail to outperform them consistently.
• Match the method to the use case.
  There is no single best model. Choose methods based on data characteristics, business context, and decision needs.
• Combine methods for robustness.
  Ensemble approaches often outperform individual models by balancing strengths and reducing risk.
• Leverage hierarchical forecasting.
  Forecast across multiple aggregation levels to balance stability at higher levels with detail at SKU level.
• Adopt AI selectively and with purpose.
  Use machine learning where data volume and complexity justify it — not as a default replacement for simpler methods.
• Account for real-world data complexity.
  Recognize that demand signals are often shaped by internal processes and may not reflect pure market demand.
• Backtest rigorously.
  Evaluate models using rolling forecasts to understand performance over time, not just in a single snapshot.
• Measure performance across levels and horizons.
  Assess accuracy and bias at different aggregation levels and across short- and long-term horizons.
• Focus on value added, not just accuracy.
  Benchmark against simple models and track improvement to ensure forecasting methods deliver real impact.

In Conclusion

There is no single “best” forecasting method, only the right combination of approaches aligned to your data, business context, and decision needs. Organizations that succeed focus not on replacing methods, but on building robust, adaptable forecasting systems.

In enterprise manufacturing, defining the right forecasting approach is only part of the challenge. Operationalizing it across complex systems and sustaining it in day-to-day planning is often even harder. Quantics is a boutique firm focused exclusively on enterprise manufacturing, combining deep data science and supply chain expertise to provide optimal support, from initial set-up to continuous optimization and operationalization of forecasting processes.

Disclaimer

In this post, we share highlights from the book “Demand Forecasting for Executives and Professionals” by Stephan Kolassa, Bahman Rostami-Tabar, and Enno Siemsen, together with our own reflections on how these ideas apply to today’s manufacturing supply chains. This is not a replacement for the book, but rather a guide to spark thought and discussion. For a deeper dive, we encourage you to explore the book itself - available here.