A first look into our Forecasting Recommender Tool
Introduction
Forecasts are of central importance in many industries. Whether it’s predicting resource consumption, estimating a company’s liquidity, or forecasting product sales in retail, forecasts are an indispensable tool for making successful decisions. Despite their importance, many forecasts still rely primarily on the prior experience and intuition of experts. This makes it difficult to automate the relevant processes, potentially scale them, and provide efficient support. Furthermore, experts may be biased due to their experiences and perspectives or may not have all the relevant information necessary for accurate predictions.
These reasons have led to the increasing importance of data-driven forecasts in recent years, and the demand for such predictions is accordingly strong.
At statworx, we have already successfully implemented a variety of projects in the field of forecasting. As a result, we have faced many challenges and become familiar with numerous industry-specific use cases. One of our internal working groups, the Forecasting Cluster, is particularly passionate about the world of forecasting and continuously develops their expertise in this area.
Based on our collected experiences, we now aim to combine them in a user-friendly tool that allows anyone to obtain initial assessments for specific forecasting use cases depending on the data and requirements. Both customers and employees should be able to use the tool quickly and easily to receive methodological recommendations. Our long-term goal is to make the tool publicly accessible. However, we are first testing it internally to optimize its functionality and usefulness. We place special emphasis on ensuring that the tool is intuitive to use and provides easily understandable outputs.
Although our Recommender Tool is still in the development phase, we would like to provide an exciting sneak peek.
Common Challenges
Model Selection
In the field of forecasting, there are various modeling approaches. We differentiate between three central approaches:
- Time Series Models
- Tree-based Models
- Deep Learning Models
There are many criteria that can be used when selecting a model. For univariate time series data with strong seasonality and trends, classical time series models such as (S)ARIMA and ETS are appropriate. On the other hand, for multivariate time series data with potentially complex relationships and large amounts of data, deep learning models are a good choice. Tree-based models like LightGBM offer greater flexibility compared to time series models, are well-suited for interpretability due to their architecture, and tend to have lower computational requirements compared to deep learning models.
Seasonality
Seasonality refers to recurring patterns in a time series that occur at regular intervals (e.g. daily, weekly, monthly, or yearly). Including seasonality in the modeling is important to capture these regular patterns and improve the accuracy of forecasts. Time series models such as SARIMA, ETS, or TBATS can explicitly account for seasonality. For tree-based models like LightGBM, seasonality can only be considered by creating corresponding features, such as dummies for relevant seasonalities. One way to explicitly account for seasonality in deep learning models is by using sine and cosine functions. It is also possible to use a deseasonalized time series. This involves removing the seasonality initially, followed by modeling on the deseasonalized time series. The resulting forecasts are then supplemented with seasonality by applying the process used for deseasonalization in reverse. However, this process adds another level of complexity, which is not always desirable.
Hierarchical Data
Especially in the retail industry, hierarchical data structures are common as products can often be represented at different levels of granularity. This frequently results in the need to create forecasts for different hierarchies that do not contradict each other. The aggregated forecasts must therefore match the disaggregated forecasts. There are various approaches to this. With top-down and bottom-up methods, forecasts are created at one level and then disaggregated or aggregated downstream. Reconciliation methods such as Optimal Reconciliation involve creating forecasts at all levels and then reconciling them to ensure consistency across all levels.
Cold Start
In a cold start, the challenge is to forecast products that have little or no historical data. In the retail industry, this usually refers to new product introductions. Since it is not possible to train a model for these products due to the lack of history, alternative approaches must be used. A classic approach to performing a cold start is to rely on expert knowledge. Experts can provide initial estimates of demand, which can serve as a starting point for forecasting. However, this approach can be highly subjective and cannot be scaled. Similarly, similar products or potential predecessor products can be referenced. Grouping of products can be done based on product categories or clustering algorithms such as K-Means. Using cross-learning models trained on many products represents a scalable option.
Recommender Concept
With our Recommender Tool, we aim to address different problem scenarios to enable the most efficient development process. It is an interactive tool where users can provide inputs based on their objectives or requirements and the characteristics of the available data. Users can also prioritize certain requirements, and the output will prioritize those accordingly. Based on these inputs, the tool generates methodological recommendations that best cover the solution requirements, depending on the available data characteristics. Currently, the outputs consist of a purely content-based representation of the recommendations, providing concrete guidelines for central topics such as model selection, pre-processing, and feature engineering. The following example provides an idea of the conceptual approach:
The output presented here is based on a real project where the implementation in R and the possibility of local interpretability were of central importance. At the same time, new products were frequently introduced, which should also be forecasted by the developed solution. To achieve this goal, several global models were trained using Catboost. Thanks to this approach, over 200 products could be included in the training. Even for newly introduced products where no historical data was available, forecasts could be generated. To ensure the interpretability of the forecasts, SHAP values were used. This made it possible to clearly explain each prediction based on the features used.
Summary
The current development is focused on creating a tool optimized for forecasting. Through its use, we aim to increase efficiency in forecasting projects. By combining gathered experience and expertise, the tool will offer guidelines for modeling, pre-processing, and feature engineering, among other topics. It will be designed to be used by both customers and employees to quickly and easily obtain estimates and methodological recommendations. An initial test version will be available soon for internal use, but the tool is ultimately intended to be made accessible to external users as well. In addition to the technical output currently in development, a less technical output will also be available. The latter will focus on the most important aspects and their associated efforts. In particular, the business perspective in the form of expected efforts and potential trade-offs between effort and benefit will be covered by this.
Image Source:
AdobeStock 83282923 – Mego-studio
