In my experience with consumer packaged goods industry, which is closest to the most fickle element of supply chain, the consumer, usage of ML provides plenty of opportunities to include demographics, retail footprint, weather patterns, CPI, employment and jobs data etc. to influence the demand appropriately. An understanding of influencing factors of demand presents great input into the planning strategy based on the strength of company, which may lie in pricing power, distribution reach, last mile presence, manufacturing flexibility etc. Even in other industries, be it B2B, there are macro trends, government regulations etc. that can act as influencing factors. With ML and its outputs, organizations can enhance demand shaping skills.

Machine learning for forecasting consumer goods is a real performance booster, in my experience at one of the world's largest retailers. Its ability to predict millions of series at a time makes it a weapon for efficiently automating forecasting, as does the ability to model complex demand patterns and take into account a multitude of variables such as promotions, prices, events, competitors, etc. Nevertheless, building an accurate machine learning model capable of modeling millions of time series is a real goldsmith's job, and requires very precise data modeling.

Challenges: - Requires large volumes of clean, normalized data. Preprocessing can be complex. - Choosing the right algorithms and architectures requires data science expertise. - Models need continuous monitoring and retraining as new data emerges. - Integrating ML predictions into planning systems requires engineering effort. Opportunities: - Uncover complex nonlinear relationships and interactions automatically. - Continuously self-improve forecasts and detect change faster. - Incorporate many data sources - sales, inquiries, search trends, weather, etc. - Personalize forecasts for customer microsegments versus aggregate views. - Allocate scarce supply more optimally when demand signals are finer-grained.

One of our key findings in 13+ years of using ML in forecasting is that it's relatively easy to train an initial model and achieve promising outcomes - what's more challenging is to sustainably deploy the model, keep it appropriately trained (features may need to be added / removed) and integrate it into the business process in a reliable & repeatable fashion. New SaaS native data platforms help with providing infrastructure to sustain such models longer term.

Data maturity, ownership, standardization, governance are some of the most touted challenges for adoption of ML. Overcoming these challenges requires strategic intent from business leadership, knowledge of business financial and operational performance drivers and linkages and technology infrastructure to make the synthesis easy and repeatable. Organizations also must develop this capability in an iterative manner to make it digestible by people, thereby enabling change adoption.

One of the biggest challenges in demand forecasting with machine learning is to model the context of our forecasts, which is made up of : Events, competitors, unknown elements from a local point of view, as well as in-store out-of-stock situations and data anomalies, all of which skew our forecasts. Another challenge we face is the lack of historical data for some products, and the intermittent nature of others, which makes series erratic and difficult to predict. At the end of the day, there's also the challenge of business adoption.

Although machine learning represents a fabulous vector for improving and refining the demand forecasting process, it is also highly correlated with the complexity of the model to be built and industrialized, and this aspect should not be minimized. So, the first step is to model this demand using very clean tabular data, without any hitches. We can then extend this huge step to a more exotic scope of data, such as text, image, video ..etc. But only if this first step is validated.

Beware of the temptation to include indices as features - a common request / idea. Most industry indices are forecasts with margins of errors in themselves. Practical reality in testing industry indices as features has often revealed that their contribution to accuracy is neglectable or negative. Better look for the underlying features that may drive an industry index and test which of those contribute to your demand picture.

Very important to not rely on a single model or method. Experiment with different forecasting models, such as time series models and also machine learning algorithms. Continuously evaluate the performance of your model using appropriate metrics. This helps identify the models that best work for the specific problem.

Among good practices, I think these are the most important: - backtest your algorithms properly, paying close attention to data leakage - don't go straight to a very complex model - build the algorithm in conjunction with the business - define the KPIs and business expectations for the algorithm from the outset - always model against a baseline

Amazon uses machine learning to forecast demand for its products. Amazon uses a variety of machine learning models to forecast demand for its products, taking into account factors such as historical sales data, customer reviews, social media data, and economic indicators. This allows Amazon to optimize its inventory levels and ensure that it has the right products in stock to meet customer demand.

Auchan Retail uses forecasting in several different ways. Here are a few examples: - promotional demand forecasting to plan supplier orders - cashier forecasts, using demand and ticket forecasts to optimize the number of cashiers allocated to each day - annual product demand forecast to place orders for new ranges

1-Consider the complexity of your data. If your data is relatively simple, a linear regression model may be sufficient. However, if your data is more complex, you may need to use a more sophisticated model, such as a decision tree or random forest. 2-Consider your forecasting needs. If you need to forecast demand for a specific product or product category, you may want to use a model that is specifically designed for that task. For example, there are machine learning models that are specifically designed for forecasting the demand for seasonal products. 3-Consider your resources. Some machine learning models can be computationally expensive to train and deploy. If you have limited resources, you may want to choose a less complex model.