Given that most supply chains operate on a demand-driven basis, accurately forecasting consumer demand has become paramount. As demand forecasting is instrumental in making data-driven decisions, supply chain managers are continually working to accurately predict future product demand. Demand Forecasting helps organizations control the efficiency of their purchasing & manufacturing operations, service levels, inventory, logistics, and pricing strategies. Overall, it has an impact on the supply chain efficiency and risk management.
Traditional forecasting approaches
Most of the traditional data-driven forecasting approaches uses historical demand data to perform time-series forecasting and estimating the future demand months ahead. Traditional forecasting involves Supervised learning techniques to build statistical models that can learns the past trends, patterns from the time series data and generates these forecasts.
The advanced industry-standard statistical models accommodate determinant variables like the seasonality, outliers and known promotional events. They are good in predicting mid to long term forecasts. These models could predict even short-term forecasts to a certain extent, while the recent consumption has no shifts or uncertainties.
Uncertainties are irresistible
However, demand forecasting has become increasingly challenging in recent years due to rapidly shifting consumer behaviors, economic factors, complex & competitive market landscape, trade wars, unpredictable events such as the COVID-19 and natural disasters or any black swan events. These factors can cause disruptions in the supply chain, making it more difficult for supply chain managers to accurately forecast future demand.
A significant challenge in demand forecasting is managing rapidly switching consumer behavior. For example, a sudden spike in demand for a particular category of goods driven by shifts in preferences/trends could deplete the complete inventory or safety stock held by distributors, leading to a more significant concern in responding quickly to replenish inventory and attain the expected levels. This is particularly challenging in modern supply chains, which are often deep-tiered and involve numerous suppliers, partners, and intermediaries, making it more difficult to coordinate and synchronize activities across the entire chain.
Managing the noise and significant abnormalities in sales history caused by long-term uncertainties and black swan events poses a significant challenge for demand forecasting too. These unpredictable events can create false trends and distort the accuracy of forecasting models, leading to incorrect predictions of future demand. Organizations are facing a bigger challenge in revisiting and tuning the supply chain strategies and policies more often than ever.
Challenges in the determinant forecasting methods
Some of the challenges in the determinant forecasting methods are (i) The traditional demand forecasting models use the historical data as the primary factor and so they can’t detect the anomalies or radical changes in the demand due to uncertainties. (ii) The models are constrained to use less variables and sources. And they need accurate data inputs. (iii) Forecast accuracy is bad in short term forecasts, while there are shifted behaviors or variations in the very recent weeks or month, which is a big challenge for FMCG companies.
Demand forecasting approaches to manage uncertainties
To avoid getting blindsided by black swan events and to manage demand volatility, organizations must implement intelligent, integrated, collaborative, and responsive processes and systems. Those include:
Leveraging demand sensing for real-time forecast
Demand sensing techniques use real-time information collection mechanisms and leverage artificial intelligence to generate more accurate & real time near-term forecasts. This enables organizations to better respond to sudden changes in consumer behaviors or market conditions, ensuring they are better prepared to manage demand volatility. By supplementing traditional forecasting methods with demand sensing techniques, organizations can gain a more comprehensive understanding of demand patterns, improve supply chain responsiveness, and ultimately improve customer satisfaction. Demand sensing systems are enabled to collect structured & unstructured real time internal and external data.
- STRUCTURED – Real-time data feeds from organization’s POS systems, Customer service, Loyalty cards, Inventory
- UNSTRUCTURED – Customer reviews from the company’s eCommerce data, Marketing Campaign data etc.
- STRUCTURED – Economic indicators, direct/indirect demand drivers, Competitor data, Customer Sales data etc., 3rd party syndicated data.
- UNSTRUCTURED – Customer Sentiment through social media forums, IOTs, videos
Real-time demand sensing leverages:
Short term forecast tuning – Demand sensing techniques uses the very recent history/trends (days or weeks) and fine tune the short-term forecast. It can accommodate wide range of factors that can affect the demand in short term, and it is capable to predict more granular forecast – like daily buckets. With the help of Deep learning algorithms (uses Artificial Neural Networks) large amounts of data can be analyzed and complex patterns can be identified – which includes the relationship between customer behavior and demand.
Centralized data warehouse – Demand Signal Repository which collects, cleanses and stores all the data in specific format for performing predictive analytics & visual analytics for demand planners. The data are not limited to promotional data, POS data, inventory turnover data, Customer Loyalty.
Text mining and sentiment analysis – Text mining and sentiment analysis are important techniques to understand customer behavior shifts and can be used to complement traditional demand forecasting methods. Text mining refers to the process of extracting useful information from unstructured data such as social media posts, customer reviews, and feedback. Sentiment analysis involves analyzing this data to determine the sentiment or attitude of customers towards a particular product, brand, or company.
Advanced data cleansing to cover the abnormalities
Before running a statistical algorithm, it is very critical in identifying the outliers and abnormal consumption patterns and cleanse the time-series data. Identifying the right pre-processing approach to adjust the historical demand for the uncertain period is vital. Some thoughts are:
- Using causal analysis techniques to identify the external factors that may have impacted demand.
- Using smoothing methods to reduce the high frequency noise.
- interpolation methods to fill in missing data points.
- Using the past or later years data, which are not/less impacted by uncertainties and override the uncertain history periods.
- Vote-out/Pruning the affected demand period.
- Using clustering algorithms to group demand patterns into similar groups and detect outliers within these groups.
- Using time-series decomposition methods to separate the data into trend, seasonal, and irregular components.
Continuously improving the forecasting models
Machine learning models used for demand forecasting require continuous improvement to ensure accuracy and reliability. This can be achieved by
- Constantly searching for new drivers and indicators to include in the forecasting model.
- Analyzing the performance of models using suitable scoring metrics such as R-Squared, Recall, and Variance.
- Assessing models with unseen data using cross-validation techniques.
- Using ensemble methods combines multiple models and improve predictive performance.
- Constantly revisit and tune model parameters regularly to improve forecasting accuracy.
By incorporating these strategies, organizations can ensure their forecasting models remain relevant and effective and are better equipped to respond to changing market conditions and consumer behavior.
Increase the Internal & External Collaboration
While forecasting models can improve accuracy, they cannot solve all problems related to the uncertainties that arise. To address this, organizations should supplement their models with information from both internal and external sources. This can be achieved by
- Overriding forecasts through expert judgements and implementing robust internal cross-functional consensus mechanisms, to ensure decision-making is based on a collective understanding of the situation.
- Additionally, real-time stakeholder collaboration (outside-in) in the organization’s supply chain ecosystem can improve the agility of the forecast to a greater extent.
For example, more frequent and granular forecast collaboration can help identify new risks and opportunities, allowing organizations to respond more quickly to changes in the market and better manage demand volatility.
Leveraging Probabilistic Forecasting for Scenario planning
Deterministic forecasting derives exact quantities, but they are uncertain. On the other side, probabilistic forecasting can accommodate uncertainty. Probabilistic forecast derives a series of probability distributions for every time periods in the horizon. It means it assigns a probability to all the possible outcomes. This approach can help organizations in scenario planning i.e. anticipate and prepare for a range of possible outcomes in uncertain conditions.
Probabilistic forecasting is well suited for supply chain scenarios like
- New-product launches
- Managing the volatility.
- Risk Management
- Predicting demands, lead times, product returns etc.
While organizations are digitizing and improving the collaboration across their supply chain tiers, incorporating the probabilistic forecasting will help the business to make informed decisions and handle volatility better.
To manage uncertainties in modern supply chains, it’s important to have agile, collaborative, and intelligent demand management systems and processes. Your organization can achieve real-time forecasting and make informed decisions by leveraging data sensing capabilities, Artificial Intelligence, Machine Learning and advanced analytics. Promoting cross-functional collaboration within organizations and across your supply chain eco system is very essential to handle the uncertain market.
Modern time-series analysis techniques like Probabilistic forecasting and models like ensemble forecasting, dynamic linear models can be particularly useful for handling uncertainties in time-series data. Your organization should constantly update forecasting models to reflect changing business conditions, identifying and addressing outliers in data sets, and fine-tuning forecast models to quantify the impact of uncertainty factors.
By adopting these strategies in the digital journey, supply chain organizations can better manage volatility and uncertainty, reduce risks, and enhance their overall agility and responsiveness.