Artificial Intelligence is Accelerating Business-to-Consumer Response

For example, if an item is out of stock, it can be quickly dropped from Amazon search results and the CPG company can lose sales. There is an emerging need for forecasting systems that can leverage machine learning (ML) and ingest Amazon ordering patterns, inventory in the channel and consumer promotions to predict the sequence of orders so that the products are in stock and the orders can be fulfilled.

In such a dynamic environment with shrinking windows to make decisions, artificial intelligence (AI) and ML techniques can enhance the forecasting of omnichannel consumer demand, shape the demand through tailored merchandising, pricing and promotions and reliably fulfill the orders for profitable growth. It is imperative for retail and CPG companies to adapt their supply chains, which were designed for a brick-and-mortar world and now need to address a wide variety of challenges introduced by e-commerce (including a CPG company’s own online channels). They continue to face challenges as network design – location and capacity of distribution centers (DC), transportation lanes and flows, SKU assortments, pricing strategies, promotion calendars that span brick-and-mortar and e-commerce, packaging well suited to e-commerce and (as mentioned above) high service expectations from retailers such as Amazon.

To succeed and thrive in such consumer-oriented markets, businesses need to become more demand-driven and orchestrate the supply chain to fulfill that demand.

Retailers and CPG companies need to predict where demand will occur, across brick-and-mortar and online channels, and efficiently supply the right quantity of products to thousands and even millions of locations.

In this journey of anticipating, shaping and responding to consumer demand, there are four core processes that that can benefit from AI/ML.

1) Forecast Demand

Forecasting the actual sell-through at the point of consumption (“Moment of Truth”, per Proctor & Gamble), whether it’s at a physical store or an e-commerce channel, drives the entire supply chain to respond. In order for the right product to be at the right location, at the right time and in the right quantity requires a complex orchestration of the end-to-end supply chain. The more accurate the sell-through forecast, the more efficient the supply chain can respond to ensure the highest customer service with the lowest investment in working capital.

AI/ML techniques can improve the accuracy of the sell-through forecast in a number of ways.

• A new set of algorithms has been shown to reduce forecast error. Support Vector Machine (SVM), Quintile Random Forest (QRF), Gradient Boosting Machine (GBM) are far better than traditional regression methods at detecting patterns and incorporating causal (internal and external) factors.
• These ML algorithms can readily leverage internal causals like everyday price, temporary price reductions, store promotions (special displays, coupons, newspaper inserts) and digital marketing programs (email offers, digital coupons).
• These new algorithms can leverage a host of external causals like weather, GDP, new housing starts, interest rates, inflation, debt to income ratios, etc. and boost the accuracy of the forecast.

Practitioners in the consumer demand forecasting discipline have found that leveraging numerous internal and external causal factors allows the ML engine to learn continuously and greatly improve forecast accuracy over time.

2) Pricing, Promotions and Markdowns

Demand shaping “what-if” scenarios can greatly benefit from ML techniques.

• A key phase in ML-based forecasting is “feature extraction” when decisions are made (algorithmically + planner input) as to which causals/features will be used as a base set to drive the algorithms. Given all the causals that could impact promotion/markdown performance, feature extraction could select just the price discount, promotion tactic and promotion duration.
• Feature engineering is another critical phase in the process which creates the proper input dataset that is compatible with the ML algorithm requirements. This is also critical in improving the performance of the ML models. For example, we may find that demand lags, promotion lags and seasonality (yearly, monthly, day of week) may be critical in improving the forecast. These engineered features also play a critical role in explaining the forecast to the planner.
• Feature/causal contribution to the forecasted volume at retail is a great benefit in planning the vehicle, timing and location of merchandising events. The merchandise planner now has the insight into how much the incremental volume can be attributed to markdown versus end-aisle display versus newspaper insert.

The merchandise planner can now experiment with “what-ifs” – look at the impact of changing the timing or duration of the markdown or combining that markdown with an end-aisle display. On the online channel, the planner can try different product placement strategies on the web site, discounts or free shipping and review the impact on forecasted orders.

3) Assortments and the Merchandising Calendar

Planning assortments across departments/store clusters and setting the seasonal merchandising calendar can be enhanced with ML algorithms.

• Which items to carry in a department in which stores has always been a challenging problem. There is a complex set of interactions between items (cannibalisation, halo, price sensitivities) and ML techniques can model these interrelationships and sensitivities.
• Feature engineering in ML can drive more accurate product life-cycle (phase-in/phase-out) planning while taking into account intra-category (cannibalisation) and inter-category (halo) effects.
• The insights from ML, like feature contributions to the forecasted volume (price, merch vehicle, placement, region, demographic), along with optimisation techniques can drive the creation of the seasonal merchandising calendar.
• ML algorithms are ideally suited for “smart” open-to-buy recommendations taking into account the merchandise financial plan, store inventory, DC inventory, in-transit, on order and vendor capacity.

4) Customer Order Fulfilment

The business-to-consumer response would be ineffective if the item the customer wants is not available or if an order promised for store pickup (BOPIS) is not ready in time. This is where the application of AI/ML techniques can be very valuable.

• Diagnostic: ML techniques can identify the root causes for fulfilment failure. There could be one or a chain of root causes that lead to fulfilment failure (such as items not in-store that have to be procured from a regional DC, workforce availability for BOPIS orders, missing items at pick locations, pickup location, pickup window duration, number/variety of items in the order, retail store space, average age of store workforce, etc.).
• Predictive: Once the root causes have been diagnosed with predictive weights for order fulfilment outcomes, the order book can be looked at and ML can then predict which ones are in jeopardy. Alerts can be sent to various roles in the fulfilment supply chain so that this can be addressed.
• Prescriptive: ML can also recommend action or a sequence of actions to the appropriate parties so that fulfilment execution is enhanced.

There are an estimated 1.5 million packages delivered per day in New York City. In the city that never sleeps, about 15% of NYC households receive a package a day. A comprehensive model of the supply chain, with predictive and prescriptive analytics powered by AI/ML, is what is emerging as the best practice in responding to that business-to-consumer challenge.

This blog was written and contributed by:

Vikram Murthi
Vice President, Industry Strategy
LLamasoft, Inc.

Originally published by LLamasoft