Forecasting Systems and Disruption with Transformers
5Summary
-------
Sathish Gangichetty, a senior solutions architect at databricks, discusses the topic of forecasting systems and how they can be disrupted using transformers. He emphasizes the importance of forecasting accuracy for companies to improve their performance and increase revenue. Gangichetty introduces the concept of transformers and their potential to improve forecasting accuracy, specifically focusing on the Patch GSD transformer. He also discusses channel independence in time series forecasting using transformers and the application of transformers in product forecasting. Gangichetty highlights the potential of transfer learning in time series forecasting and the importance of creating user-friendly interfaces for forecasting systems. He concludes by discussing the simplicity and accessibility of surfacing results using transformer models and the need for trust in the user experience. The presentation also briefly touches on the challenges of stock prediction using reinforcement learning models.
Topics:
-------
Introduction
* Sathish Gangichetty is a senior solutions architect at databricks
* Forecasting accuracy is crucial for companies to improve performance and increase revenue
Challenges of Forecasting
* Forecasting is a difficult problem to solve due to its reliance on time-based data
* Unpredictable events can impact forecasts
* A 10-20% improvement in forecasting accuracy can lead to a 3% growth in revenue
Transformers in Forecasting
* Patch GSD transformer challenges the notion that transformers are not suitable for time series forecasting
* Channel independence in time series forecasting using transformers improves training results
* Patching enhances local semantic information in the transformer
Application of Transformers in Product Forecasting
* Transformers can incorporate multimodal models for more accurate estimation of demand for new products
* Transfer learning can be used in time series forecasting to forecast new data without additional training
* Creating user-friendly interfaces for forecasting systems improves usability
Surfacing Results using Transformer Models
* Transformer models ensure privacy and determinism of results
* Anyone can interact with data and utilize these models without being a specialized data professional
* Trust in the process is created through consistent results and user-friendly interfaces
Challenges of Stock Prediction
* Stock prediction is a difficult task even with advanced models
* Unforeseen events can greatly impact stock prices, making it challenging to predict with certainty
* Flexibility and the ability to react quickly are important in the stock market