Price Elasticity Key Takeaways:
- There are multiple shortcomings to the traditional method of price elasticity analysis
- Virtual shopping has many advantages for identifying own and cross price elasticity
- Machine learning algorithms simulate tens of thousands of scenarios to estimate sensitivity of elasticity in real-time
Webinar Hosts 🏷️
Dr. Kurt Jetta, Chief Analytics Officer, TABS Analytics
Jordan Henderson, VP Client Solutions, Decision Insight
Scott Tranter, CEO, 0ptimus Analytics
Dr. Kiel Williams, Data Scientist, 0ptimus Analytics
Price Elasticity: The Traditional Approach
Elasticity measures the change in demand when economic factors (like price) are adjusted. For CPGs, elasticity is a key concept in improving pricing strategy.
Elastic = A change in price results in a change of demand.
Examples: Luxury items, commodity products
Inelastic = No change in demand even if prices changes.
Examples: Gas, medications
Traditionally, you look at price and quantity to understand elasticity.
In Chart A (below), look at the orange line, that’s the fitted curve. The yellow line demonstrates that there’s a relationship when the price drops, the estimated baseline goes up. This only focuses on everyday pricing – not promotional pricing.
In Chart B (below), look at the other factors that could impact estimated sales. Ask, What’s the price / quantity relationship? If the relationship exists, you should see it in the data.
This is all we could ever hope for in a price elasticity analysis. Great fit – great relationship. But, this approach isn’t going to work anymore…
Shortcomings of Traditional Price Elasticity
(video time: 3:50)
The catalyst for rethinking price elasticity, is COVID.
In the graph below, look at the baseline (purple line at the bottom), you see where the COVID bump occurred in this product. Even within that bump, there’s different demand regimes – demand goes up, stabilizes, then goes down.
Based on preliminary data, it appears that consumers (during the COVID bump) became less price sensitive. The overall demand went up 7% and average pricing in the CPG industry went up 7%.
This shift in base volume means we have a whole year of data we can’t even use!
Another problem with the traditional price elasticity analysis approach, in the US, “despite what everyone tells you about store level pricing and consumer demand, it’s not true. The majority of the time, there is chainwide pricing that stays consistent for an extended period of time, and we can’t use most of our data sets because there’s no price variation,” says Dr. Kurt Jetta.
You can see that in the graph on the right (in the image above). This particular product, the price didn’t go up during COVID. Yet, you see those promotional pricing and promotional lifts are off the charts.
(video time: 6:45)
So many people want to know cross elasticity effects versus their own. My elasticity means if I take a price change, what’s the impact on my volume? Most CPGs want to say, if my competitor changes their price, what happens to my sales?
Cross elasticity of demand measures the percentage of change of the quantity demanded for a good to the percentage change in the price of another good.
Often, we think about elasticity as 1:1. There’s Product A and Product B – the relationship is positively sloped cross elasticity – which means these products could be substitutes. But, cross elasticity impacts multiple products and segments.
EXAMPLE | Category with 10 brands… 50 brand sizes… across 30 geographies… = 73,500 cross elasticities to analyze
That’s an unmanageable number of analytics to conduct with in market scan data.
A further problem: When you do get some type of movement in pricing, close competitors match. That blows away any ability to get an isolated analysis on Brand X takes price up, what’s the impact on Brand Y… You can’t conduct cross elasticity.
(video time: 9:25)
Price elasticity is a marketing mix model. “We still need to control for all the key marketing variables. It becomes a complex model,” says Jetta. “The more variables you add, the more likely you’re looking at collinearity… which is another shortcoming.”
- Consumer Promotion
- Marketing Communications
- Market Shocks
(video time: 10:55)
Cost & Timing
To turnaround a price elasticity analysis, it takes 6-8 weeks, costs about $60,000 – $75,000 per category, and is analyzed once per year.
Then you have to rely on that all year, even if things change. Wouldn’t you prefer to read this analysis as an ongoing deliverable?
(video time: 11:57)
As an industry, what we’re doing now “is not adequate,” Jetta continues. That’s why a whole new approach is warranted. There has to be a better way to get quicker results, more dynamic results, have more granularity, and be able to easily respond to any changes.
NEW Approach = Virtual Shopping + Machine Learning
Virtual shopping is a predictive tool that helps CPGs figure out how shoppers will act in the market. Virtual shopping can happen on your laptop, smartphone, or PC; it can happen across retailers, geographies, and products.
5 Benefits of Virtual Shopping
- Measures actual sales (same as in-market test)
- Fast, current market response (based on current market conditions)
- Forward looking with regular updates (weekly or monthly)
- Reduces risk (isolates test variables for targeted cross elasticity effects)
- Timing and cost (great efficiencies)
(video time: 15:25)
“At Decision Insight, we’ve been doing virtual shopping research for more than 15 years,” explains Jordan Henderson. “We regularly take our sales data from virtual shopping and match it back to in-market data and we’ve found that it correlates at .9 or higher.” The data that supports this shows that it works. “We know that what you’re measuring is what’s actually happening in the marketplace.”
Virtual shopping research correlates at .90 or higher.
Virtual research can handle compounded effects – you can layer in lots of variables.
There’s a wide variety of what you can test when it comes to virtual shopping and machine learning.
Respondents go through the virtual shop test
…that data feeds into the analytical engine
…which is powered by machine learning
… you get the results.
(video time: 17:51)
Machine learning models trained on historical consumer data are used to simulate cross-elasticity effects on demand within the marketplace.
Takeaway: The same amount of time, the same amount of resources, but an increased amount of insight.
You could have a very simple linear regression approach to understand market dynamics. You might even have something like that in Excel. It’s very limited in terms of market behavior in what you can understand and analyze. Imagine going to something more sophisticated, a one pipeline or single machine learning approach. Dr. Kiel Williams explains, “We take the next step. We take an ensemble approach that combines a variety of different machine learning approaches. You get a richer, fuller picture of how different products within a given market are interacting with one another.”
Given all the virtual shopping scenarios, this ensemble machine learning approach can plow through thousands of scenarios to simulate price elasticity analyses.
(video time: 19:50)
In the Chart C (below), there’s a comparison of salty snacks data.
Advanced machine learning ensemble modeling outperforms linear regression with an average R2 20% higher across all brands without sacrificing time or speed.
Because of computing power, we can pull out the best insights.
Williams continues, “To comment on this bar graph that’s comparing R2s, the linear regression approach in green and compare it to the the machine learning models in orange – you can see that the machine learning model consistently provides a more accurate picture of how these markets are interacting than you can get from that simplistic linear regression approach.”
On the backend, there are a few thousand lines of code. There’s lots of multithreaded processing and data handling. On a normal desktop, you might be able to turnaround the data in a week. But, 0ptimus Analytics machine learning tools can turn the data around in a few minutes. The speed is one thing, but the ability to capture non-linear relationships is even more important. It’s an increased look at how products interact and where demand transfer comes in.
(video time: 22:41)
When we think about cross elasticity, it’s fundamentally non-linear, Product A demand changes as price for Product B changes. If you increase a product price by $1 versus $2, the magnitude of that effect on a different product is not the same at each of those price points. By construction that’s impossible to pick up via linear regression. But, through a machine learning approach, that data is captured.
“Stats are nice, but if the data doesn’t help you change your strategy or confirm what you’re doing, then it’s all vanity,” says Scott Tranter of 0ptimus Analytics. Advanced machine learning algorithms are used to extract patterns in consumer behavior. Tens of thousands of simulations are generated to explore the possible pricing scenarios. Simulations are used to estimate the sensitivity of cross-elasticity on individual products.
DEMO: Cross Elasticity Demand of Salty Snacks (video time: 26:16)
With static simulation, you’re able to explore any set of product prices within the market assuming competitors don’t respond at all.
For some product categories, where prices change slowly, this is appropriate. For faster moving categories, you can gain insights, but you also want those agent-based simulations.
In this example (image below), let’s look at salty snacks it the Latin American market. Ruffles Queso, Doritos Nacho, Cheetos Puffs, etc. Let’s look at the base case inside this market; the simulation shows our own portfolio. The products are indexed to 100. (In the market, if demand is indexed to 100, it’s more popular than the average.)
That black line within the index is the confidence interval. There’s no delta yet, because we haven’t changed any prices around the simulations. You can see competitors too.
Let’s change the price of a few of our own products. We’ll change pricing for Cheetos products from $8 to $10. Time for ensemble modeling!
We’re going to assume competitor prices stay the same.
As expected, because we raised the price on Cheetos, demand went down, but our Ruffles Queso chip demand went up.
The model shows own portfolio value will go down 1% but the overall market will stay the same.
This price elasticity methodology shows that multiple products are impacted by price changes.
Through this methodology, you see if prices go up…
- Demand will go down (at least in this category / scenario)
- Some people won’t buy any chips
- Most people still buy chips, just buy a different chip (which kind of chips will they buy instead)
- Explicitly characterize demand transfer
(video time: 31:15)
With agent-based dynamic simulations, you change your price, and the model determines how competitors might respond. Say you increase your price, it’s likely your competitor increases its price as well, and in that particular framework, how does demand for all of the category change as a response?
If we change our prices, now we’re assuming that the competitors changed their prices too. You can set thresholds. Identify the base case… our own portfolio is worth $7,700. We’ll move the price (again) from $8 to $10. The tool runs scenarios and anticipates what your competitors might do.
Even when we increase the price of Cheetos, because of what our competitor will do, our portfolio sees an increase of 3%. We don’t see a decline. The overall market grows by 4%.
Because of the wide product base, not every shopper flows to the competitors.
This is a powerful approach, especially in this COVID-demand, lets you test with reliability.
(video time: 35:45)
Traditional elasticity analysis is inadequate for CPG brands over the next 2-3 years. That COVID bump creates a problem.
As a replacement, we have virtual shopping that correlates with in-market results and has numerous advantages for identifying own and cross price elasticity.
The insights of virtual shopping can’t come alive unless machine learning is utilized to cull through tens of thousands of pricing scenarios. But, no need to overengineer if potential combinations are <10,000.
For highest confidence, virtual shopping / machine learning results should be corroborated with in-market results from traditional price elasticity analysis.
With a traditional conjoint – you’re limited in scenarios. With machine learning – it’s possible to understand scenarios without any specific conditions.