Whether your focus is on energy loads, production process schedules, or inventory forecasting, one thing remains certain: effective time series forecasting requires far more than simply selecting an algorithm. Historically, identifying the optimal model demanded not only a robust data foundation but also exhaustive experimentation and constant fine-tuning.
At SOTEC, we have changed this game for our customers completely.
Forecasters Race
Modern data ecosystems have commoditized complex model building, drastically lowering the barrier to entry. Instead of agonizing over the perfect architecture, we now rapidly benchmark a suite of candidates, in the so-called “forecasters race”.
The result? Less time tuning parameters, and more time focusing on the domain insights and operational decisions that pushes our customers’ businesses forward.
But: This blogpost shall not be about the details of the forecasters race. Check out the excellent article at towardsai.net from our Google-fellow Tuba Islam, if you want to know more about the methodology.
SOTECs take – AI Forecasting Ecosystem
The best forecasting approach is useless if there is no data foundation to work with. So, we at SOTEC embedded the forecasters race approach into our solution for data-driven processes, therefore inventing the AI-forecasting ecosystem:
- Multi-source Data Analytics: You get to acquire the data at the source and the source is often on the Edge. SOTEC is a pioneer for many years in field data acquisition, from Industrial IoT to Retail POS systems.
- Data Warehousing: Data is centrally stored, organized, and maintained, with BigQuery being our designated solution.
- Data Engineering: This involves cleaning, curating, and augmenting data, including the harvesting of exogenous data (e.g., economic indexes, port/freight status, web trends, weather data).
- The Forecasters Race: This is the process of testing various forecaster architectures to identify the best performer (as brilliantly explained by Tuba). We enthusiastically let race all available GCP AI models—from lightningspeed Deep Neural Network training on BQ ML and tailored LSTM on Vertex AI to Zero-Shot TimesFM—to ensure multiple strong candidates, diversified strengths, and reliable backups.
- The Learning from the Learning (Explainability): Moving beyond simple data analytics, we use explainability features to understand backward what drives the forecasts. Analyzing the forecast drivers unlocks deep insights and improvement opportunities.
- Serve the Forecasts (Operationalization Cornerstone): Building and deploying services that deliver forecasts precisely where and when they are needed. This is critical for operationalization and involves managing critical factors like forecast latency, operational cost, automatic model retraining, and scheduled data engineering backend jobs.
- Visualize: The true value of a forecast is realized when it reaches the end user—be it a logistic center operator, a store manager, or a data analyst—via the most efficient medium (e.g., a mobile app, a Looker dashboard). We place strong emphasis on presenting forecast information to users effectively.
At SOTEC, we invented this ecosystem to tackle all aspects of what makes data-based insights, decision-finding precise and therefore valuable. The embedding of the forecasters race approach was the perfect addition to it, enabling us to build solutions with the capability to forecast what “will probably be”.
Keep an eye out for our upcoming blog post, where we will dive into a real-world application of our AI-forecasting ecosystem to tackle dynamic localized forecasting on global scale.