The Gist

  • Data importance. AI and predictive analytics require high-quality, diverse data sources for effective ecommerce applications.
  • Customer insights. These tools enable businesses to better understand customer behavior, forecast demand and optimize pricing strategies.
  • Continual improvement. Regular monitoring, testing and updating of predictive models are essential for maintaining accuracy and achieving desired results.

Artificial intelligence (AI) and predictive analytics can have a deep impact on ecommerce across different functions and processes, including demand forecasting, inventory management and sales forecasting as well as pricing and promotions.

These two tools are essential to predict customer behavior and forecast demand, analyzing large amounts of customer data and identifying the patterns and trends held within that data.

By using AI to identify trends and patterns in consumer behavior, online retailers can better forecast demand for products and set prices that maximize profitability.

Organizations can also glean insight from historical data to predict future customer behavior and the products and services they are likely to purchase again, how much they are willing to pay and when they are likely to purchase.

Pulling Data From Multiple Sources in Predictive Analytics

Bill Nowacki, KPMG managing director of data and analytics, explained a variety of sources are often used to inform forecasting, purchasing, pricing and other related planning exercises, including those describing consumer spending or costs of input — such as labor, energy and transportation.

"Adequate options for each of these categories exist within the public domain with many of the publishers providing APIs," he said. "Ecommerce companies can engineer pipelines to pull data directly from the sources."

These data sets are often geotagged and accumulated into time series, then examined through feature engineering to determine the degree to which signals might be useful in determining or predicting local, regional, or national demand or price sensitivity. 

For ecommerce businesses considering implementing predictive analytics and AI for their inventory pricing and sales strategies, Nowacki recommended first assembling a robust data ecosystem — the bigger the better.

"Second, use machine learning to determine the drivers and indicators of demand for the company’s products or services," he said. "Endeavor to understand demand and the impact of pricing on demand at both the hyperlocal and national level."

He added it's important to recognize that there are many dynamics that affect commerce. The more features one can engineer into the models, the better, and the more extensive the use of exogenous data, the better. 

Related Article: The 5 Stages of Predictive Analytics for CX Success

Using AI to Improve Pricing, Sales Decisions

Darren Saumur, president of cloud and customer experience with Blue Yonder, said when it comes to pricing, AI and predictive analytics provide visibility into how far you need to adjust your prices to incentivize customers since that is a personal decision.

"There is a fine line between providing offers that drive demand but destroy margins," he said. "Getting that right can save a lot of margin."

He noted AI has a significant role to play in competitive head-core-tail pricing.

By using historical sales, price sensitivity and the associated margin impact, the products can be classified into head, core and tail with different pricing strategies. For instance, the lowest price strategy may be applied to specific product groups.

Saumur added AI can help determine how and where to place inventory geographically, but what's critical is having the right customer data.

"Looking at all data along the customer journey from browsing online products, to those they buy, to what gets returned, as well as which promotions drive the most traffic, are some of the data that help improve the accuracy of the predictions on pricing and sales," he said. 

Learning Opportunities

He points out demographics and other meta data about your online customer can also help target sales campaigns and offers.

Related Article: Predictive Analytics: Overcoming Data Swamps in Tech's Dynamic Landscape

Ensuring Quality Data for Better Predictive Analytics

Forrester principal analyst Biswajeet Mahapatra explained the key to maintaining accuracy of any model is the quality of data.

"Having a robust data governance process in place with a proper documented data strategy leads to better data quality," he said. "Data quality can be insured with periodic data testing and cleaning."

Another important aspect to maintain the accuracy of predictive models is usage of appropriate algorithms and regular updating of the models.

"Regular monitoring, testing and updating of models is essential for accurate results," Mahapatra said. 

He cautions implementation time can be long and tedious as it depends on a lot of factors, including data availability, data quality, data mapping and integration, tools deployment, model deployment and testing and skill set availability.

"Understand and document your objective of implementing predictive analytics and AI and what is the end goal you want to achieve," he advised. "Without having a documented end goal and target in mind you will not be able to measure your progress and success."

He added ecommerce companies will often perform A|B testing, for example experimenting with price or other sales considerations in a comparable market to witness and measure the effect of the experiment on product sell-through.

"Always compare 'actuals' to 'forecasts' to ensure that models are solid and to also determine if the underlying market assumptions have changed," he said. 

Saumur cautions that overtrusting models can also be a potential drawback, which is why businesses must test small segments.

"One of the drawbacks with anything statistically based is the propensity to overfit models to results; you train your model to deliver the results that you expect," he said. "Then you test the model that delivers the results you want. This is why it’s important to go back and continually do micro-testing to ensure that your model is as bias-free as possible."