Inspiration

Sales forecasting has always been an important and difficult problem to solve. Accurate and reasonable results are essential to allow firms and companies to manage their inventories smartly and boost revenues by saving on logistic costs and running an efficient supply chain mechanism. With recent advancements in AI methodologies, it is now possible to build models that are specifically tuned to capture trends in data depending on a variety of external aggressors.

What it does

We have a cutting edge AI approach to try to predict trends in sales data. Given a dataset of historical sales data, our approaches learn trends in the data on a location and product wise basis and predict future sale numbers and sale trends for a given product in a given locality. We compared our approach to the best time series predictor currently in the field, namely Prophet by Facebook, to ensure results given by our model are reliable and accurate.

How we built it

Deep learning methods are employed with a fully connect neural network with drop out and batch normalization layers to prevent overfitting for small datasets and to ensure the model architecture remains complex enough to capture non-linearities in data allowing it to learn trends in the data. The model was trained on two different datasets showing good results on both.

Challenges we ran into

Acquiring a dataset that had enough data to train a deep learning model and one with information regarding multiple locations and multiple products was a significant challenge since deep learning requires much more data than a conventional machine learning approach. Integration of layers into the model and avoiding overfitting while still scaling the model to be more accurate was also a significant issue we had to overcome.

Accomplishments that we're proud of

Given our primary dataset, our model has learnt trends for items very well with a mean absolute error in the single digits for almost the entirety of the products and some scores ranging as little as between 3 and 4. Given the average value of sales in the dataset varies from about 25 to over 50 for different items in the data, this is superb accuracy comparable to the accuracy of Facebook's inhouse architecture for time series forecasting.

What's next for Sales and Inventory Forecasting using Deep Learning

Better datasets and more optimized models given more time may improve this accuracy further. Moreover, models tuned towards event detection may be used to detect breakouts in the sales data which when used in conjunction with the trend model may further increase the accuracy to include more outliers in the data making the model even more accurate.

Built With

Share this project:

Updates