When it comes to boosting sales, fresh department managers face a balancing act between maintaining fully stocked shelves and minimizing shrink. Staffing shortages and the rise of online grocery shopping further complicate this act, however, the root of the problem lies much deeper. 

Perpetual inventory (PI) systems fail to track fresh inventory accurately because they don’t account for fresh-specific variables like spoilage and unrecorded shrink. To address these unique challenges and deliver greater value for grocery retailers, Afresh developed a new inventory estimator, InvHMM (patent pending). 

InvHMM – which stands for Inventory Hidden Markov Model – leverages artificial intelligence and machine learning to offer a more accurate, labor-efficient method for inventory management, addressing the unique dynamics of fresh where PI systems fall short.

This blog offers a peek behind the curtain to understand how Afresh is purpose-built to help grocery retailers stand out with well-stocked fresh departments featuring the highest-quality products. But first, let’s explore the pitfalls of PI-based solutions.

The Pitfalls of Perpetual Inventory

Perpetual inventory (PI) systems are a traditional retail tool for tracking stock. They calculate current inventory by adding incoming goods to – and subtracting sold or lost items from – the last known stock count.

Unfortunately, these systems fall short in the dynamic world of fresh due to unique challenges like:

• Spoilage,
• Unrecorded shrink, 
• Weight variability,
• Labor-intensive tracking processes that increase human error,
• And the absence of real-time data. 

Without an accurate inventory count, fresh departments are prone to overstocking (thereby, increasing shrink) and understocking (and missing sales opportunities) – both of which lead to reduced profit margins. To compound the issue, inefficient inventory management can also reduce the shelf life of fresh products, negatively impacting the overall customer experience.

The Impact of Machine Learning on Fresh

‍Hidden Markov models (HMM) are statistical models used in machine learning to estimate the impact of “hidden” (imprecise, unmeasured, and dynamic) variables in uncertain scenarios. One common application of this is spelling autocorrect – though your phone cannot know exactly which word you’re attempting to type, HMMs allow it to consider the context of the surrounding words as well as your past mistakes and writing style to make an educated guess.

In fresh grocery, we know that inventory is inherently variable and subject to unobserved factors such as unrecorded shrink – HMMs allow us to adapt to this uncertainty to better estimate inventory. 

At Afresh, we apply InvHMM (our inventory estimator powered by hidden Markov models) to several key areas to improve inventory measurement:

1. Higher accuracy inventory estimation: Afresh’s probabilistic approach to inventory drives higher accuracy by running hundreds of simulations to model the potential impact of uncertain variables such as unrecorded shrink on the observed inventory level of a given item.

These simulations return a range of potential inventory trajectories (i.e. series of inventory positions over time), each weighted by the likelihood of seeing a given observed inventory on a given day if a specific estimated previous inventory trajectory were in fact true. The models continuously compare these weighted estimates to the observed inventory levels over time, recalibrating based on incoming data, model assumptions, and input from store teams when necessary to avoid the cumulative error we often see with perpetual inventory. The result is a more accurate and resilient reflection of a given item’s inventory position over time.

For example, if Monday’s observed inventory is 10 units and a shipment arrives Tuesday with 5 units, the actual inventory for Tuesday could be any value between 0 and 15. The weighted probability of each potential inventory value will vary based on forecasted sales, shipment fill rates, and perishability/shrink.

Assume you extend the InvHMM model to predict inventory on Wednesday with an expected incoming shipment of 7 units; as such, the possible values for inventory could now be any value between 0 and 22, again with different weightings across each value. 

However, if you observe 6 units of inventory on Tuesday (all 5 units of the shipments arrived, 8 units were sold and 1 shrunk out), InvHMM would adjust the range of potential values and weightings to reflect that it is now highly unlikely that the actual inventory level is 22 -- recalibrating uncertainty levels based on newly-observed data.

2. Resilience to human error: PI-based systems are particularly vulnerable to human error, as they snap to the most recent inventory measurement and take any given input as truth. In contrast, InvHMM tests each input against historical inventory levels and sales / shipments data to identify potential inventory measurement error. Afresh can also prompt the user to correct likely errors, ensuring the system continues to get accurate reads of observed inventory to calibrate its estimates.

For example, assume it is Monday in store again and you observe 10 units on hand. If on Tuesday, you believe stores received a shipment of 5 units and sold 7 units since the inventory count on Monday, a PI system would anticipate the inventory to be 10 + 5 - 7 = 8 units. If a store inputs 7 units, this could be likely because 1 unit shrunk out – this feels feasible given our understanding of potential unobserved factors.

However, if a store inputs 3 units, Afresh might prompt the user to double check as this would imply that 5 units shrunk out, which is more than expected. This inventory alert would prompt the store team to double check; perhaps the store missed counting the 5 units that arrived on Tuesday as they were tucked in an unusual spot in the back room.

3. Targeted store team engagement: InvHMM also detects where uncertainty in inventory data could lead to an inaccurate order or stockout in order to focus store teams’ time exclusively on the highest-value activities.

For example, if Afresh knows the likely range of actual inventory levels for cashews is 30-40 packs and that upcoming demand is likely 10-20 packs, it would recommend an order of zero no matter what – store teams do not need to review this item. However, if upcoming demand were 25-35 packs, Afresh would detect a potential stockout risk (e.g. if they actually had 30 packs but demand ended up being 35 packs) and prompt a store user to input inventory to increase the precision of our inventory estimate and, by extension, our order recommendation.

4. Optimize merchandising standards: Lastly, the probabilistic inventory estimates produced by InvHMM make it possible to more accurately model expected shelf fullness and stockout risks over time. This opens the door to providing greater visibility to merchandising teams as they seek to optimize product availability and customer experience. 

The Top 4 Benefits of Afresh’s InvHMM

When we apply InvHMM to real customer data, we find that our model improves accuracy over perpetual inventory (PI) between 10–40%. But accurate inventory is just the beginning. Here are the top four benefits of Afresh’s InvHMM:

1. Increased Accuracy & Robustness to Errors: By leveraging probabilistic models, Afresh's system vastly improves the precision of inventory estimates, effectively minimizing common errors that plague PI-based solutions.
2. Quantifying Uncertainty: Unlike PI-based systems, Afresh quantifies inventory uncertainty, providing retailers with actionable insights to make informed decisions and optimize stock levels.
3. Greater Labor Efficiency: Afresh significantly reduces the number of manual inventory checks, thereby saving valuable labor time and resources.
4. Reduced Shrink: By placing truck-to-shelf orders based on more accurate inventory counts, Afresh's system effectively reduces shrink, lowering waste and boosting profitability for retailers.

These benefits collectively demonstrate how Afresh's innovative approach to inventory management addresses the dynamic nature of fresh department operations, leading to improved operational efficiency and financial performance​​.

Ready to transform your fresh department?

In summary, Afresh’s InvHMM marries the predictive power of machine learning with the flexibility of hidden Markov models to offer grocery retailers a robust solution that not only enhances inventory accuracy but also optimizes ordering and mitigates shrink. 

More importantly, InvHMM is designed to empower store teams. By pinpointing where manual interventions are most impactful, InvHMM ensures that ordering decisions are based on the most accurate data, every time. In this way, Afresh strikes the perfect balance between technological efficiency and human expertise.

Schedule a demo today and learn how Afresh’s InvHMM can help increase efficiency, drive sales, and reduce shrink across your fresh departments.