Anisha Dubhashi
Data Scientist - Nordstrom
Anisha is a Data Scientist on Nordstrom's Merchandising Analytics team. She works on developing analytics products related to inventory forecasting, price optimization, and size curves. She holds a master's in analytics from Northwestern University and a bachelor's in mathematics and economics from UCLA.
Dana Lindquist, PhD
Sr Technical Program Manager - Nordstrom
Dana discovered data science about 4 years ago while working as a project manager at a company that had a great deal of data. Her career had trended away from numerical methods where she received a PhD some years ago. To get more involved with data science she enrolled in the Metis Data Science Bootcamp after which she worked as a data scientist for 3 years. She recently joined Nordstrom as a Sr. Technical Program Manager for a data science team, pulling together much of her past experience.
Jessica Marx
Data Scientist - Nordstrom
Jessica is a Data Scientist at Nordstrom where she builds analytical tools and software supporting Merchandising and Price; past work includes supporting Clickstream and Product. In addition to Nordycast, she's worked on Assortment Health/Optimization, Duplication Metrics, and Markdown Optimization. Through internal programs she teaches SQL and Python to stakeholders and junior team members. She has an undergraduate degree in Film from NYU (how she got from there to here is a long story). If it's sunny outside, she's probably roller skating.
Nordycast: A Flexible Forecasting Library
Session - Anisha Dubhashi, Dana Lindquist & Jessica Marx
There are many teams and countless individuals creating and using forecasting models at Nordstrom. This represents a huge duplication of effort, but until now there was no effective way to efficiently share these models and the code surrounding the models. Enter Nordycast, a way for novices to use existing models and experts to share the models they have created. Nordycast aggregates all the setup and visualization code that is needed when creating forecasting models and is focused specifically on forecasts using Nordstrom data. It is the company’s first venture creating a shared library. High-quality forecast modeling is hard to do and time-consuming. It starts with pulling together the relevant data needed for the forecast, setting up target variables, adding lags and seasonality, filling in missing data, and creating a train-test split. Only when this data preparation work is done can the actual modeling start. Forecast modeling steps generally include hyperparameter tuning using time series cross-validation, creating a naïve baseline model for model evaluation, and aggregating multiple models to create an ensemble model. The final models and corresponding metrics need to be saved to be used in the future, and the entire pipeline may need to run in parallel depending on how many models exist. During this process, the features, model metrics, and output need to be visualized. Nordycast simplifies this entire process by providing functionality to prepare the data, create multiple forecasting models, parallelize the pipeline, and visualize the results. Nordycast is also tailored to the type of forecasts and data used at Nordstrom. Since Nordycast is an open-source library (inside the company), data scientists can contribute new models and new functionality to the library. Experienced data scientists will be able to add to the library of models, giving them a place to highlight and share their work. Less experienced data analysts and data scientists will be able to quickly spin up a forecasting model. The beauty is, that the repetitive work of creating a forecasting model is packaged in Nordycast, allowing someone creating a new model to focus on the model. To date, the response at Nordstrom has been very favorable. Users of Nordycast are particularly excited about the ease of using Nordycast and the significant time savings, especially in model setup. Future contributors are also excited about adding deep learning and time series forecasting methods to Nordycast’s model library. In our talk, we will review our process of building Nordycast, maintaining the open-source package, user engagement, and our plans for the future.
