Feedback Loops
Learning how to model based on Spotify’s business model
Overview
Have you ever been in a meeting and someone proposed a plan that didn’t take into account user behavior? I have been in countless meetings where a decision was made without considering the way the system would work. It is not because someone didn’t intend to think of the unintended consequences. It is likely because they didn’t take the time and/or know how to model the feedback.
As an engineer I was trained to design feedback loops in a system, though to apply this in business has been harder. For one, you are dealing with people (aka irrational actors) and two the idea of modeling business as a system is still in its early years — well technically Jay Forrester of MIT in the 1950’s was the first to start using the concept. My first exposure was when I was looking at business schools and audited a Systems Dynamics course at MIT. From that single course I was hooked. It was the intersection of my engineering degree and my future goals in business.
In this post I’ll share the basic concepts of system dynamics and an example to help you start to think about how you would identify behaviors in the business world. The goal of system dynamics / system modeling is improve your understanding for how your business works.
Key Concepts
I’m going to focus on a few items that will help to model out feedback loops. If you are interested in more I have included some reference materials below.
- Reinforcing Loop — This is a positive feedback loop where you get more of the result. E.g. If people lived forever we’d end up with more and more people
- Balancing Loop — This is a negative feedback loop where eventually you end up at zero. E.g. A really high mortality rate would drive the population to zero.
- Behavior over Time Chart — This is the output of the feedback loop in data, which is useful to understand the impacts. Below are a few examples based on various scenarios one may see.
With these concepts in hand let’s start with a few easy examples to understand the concepts and the structure of a simple model.
Example — Spotify’s Growth Model
In this example we’ll base it on growing adoption of a new product that you are taking to market. For this example let’s assume it a subscription service, and given Spotify’s recent S1 announcement to go public we’ll use it as the basis for analysis. The goal of the analysis will be to determine how the Spotify Profit Management system works. To start we’ll identify the dominant feedback loops that have direct financial impacts. There are other loops that I did not include like more content and improved recommendation algorithms that improve the product experience.
What are the reinforcing loops?
- More product usage= higher retention
- More marketing money spent = more customers (and awareness)
What are the balancing loops?
- Price increases = reduced retention and user signup
- More adopters = fewer potential subscribers
Behavior Over Time Charts
Now that we have the basic feedback loops defined, let’s start to build out our view of the system. In the images below I will show a visual depiction of each of the reinforcing and balancing loops. In addition, I’ll add in a Behavior Over Time (BOT) chart. These charts depict the behavior that is seen from the feedback loop.
Examples of Reinforcing Loops
Building upon the two loops we defined above, let’s look at the causal loop and the BOT charts.
For the first feedback loop of product usage = higher retention, I have modeled out that more usage of the product has resulted in higher retention data. It makes sense that people who use a subscription product the most are more likely to stay around. In the balancing loops we’ll look at what would cause someone to leave.
For the second chart, I have modeled out two different scenarios for the future based on their current growth rate percent and growing their net adds at a constant growth rate. The difference is very stark, with the Growth Rate model resulting in 75% more users in 5-years. In evaluating their data it is more likely that the net add approach is a more realistic view, but both are reasonable projections.
I also looked at the customer acquisition costs (CAC) by quarter and found a slightly increasing trend in the costs to acquire subscribers. This would be a metric to monitor over time to understand if it costs more to acquire additional subscribers. This over time could have a large impact on their economics, especially given that marketing has risen from 10% to 15% of revenue over the past two years.
Examples of Balancing Loops
The two balancing loops we defined above are depicted below over a 5-year time horizon.
The first one on price shows the current churn rate on users (which I’m assuming for simplicity is a function price point). For Spotify they are at a 5.5% churn rate, which is an improvement from 7.7% two years prior. Based on the S1 this was due to family and student plans being introduced to target what were likely high churn segments and growth opportunities. The current rate would imply that within 5 years nearly all current customers would have churned and after 1-year Spotify loses 40% of its paying base. While Spotify added 23mm users from 48mm to 71mm in 2017, to get there they lost ~36mm users and thus added 62mm total users (or nearly 3x the net adds).
The second graph above shows how the potential market will shrink as they grow their base, using the projected growth rates of the business. The orange link represents the interaction with their churn rate. Assuming their current churn rate and acquisition rates they will have served 500mm users within 5-years. This would suggest that they will need to be able to serve a broader audience and reduce churn rates.
What are the delays that we should consider?
One of the additional mechanisms to evaluate is whether there are any built in delays in the system. An example would be pregnancy takes 9 months, so you would need to factor in this time delay. For the Spotify example they have recently started promoting a 12-month plan at a discounted rate to the monthly premium plan. There wasn’t any specific data, but here is how I would represent that in a causal diagram:
The idea of having a contract is not without contention. The wireless carriers in the US have been battling this one for years. In my experience, if you have a good product you will see the economics are very similar for the two models. The best products in the world don’t need a 12-month lock-in.
Integrated Model of the System
I have put together an illustrative (ie. quick diagram) to portray many of the factors above into an integrated system view.
With this model I then took a look at the BOT chart for the company (see below). Based upon the above feedback loops we discussed it would make sense that Subscribers (users) is growing faster than revenue due to lower pricing options. The Cost of Revenue increasing at a slower rate than revenue is a huge factor for this business and the results of these savings helps Spotify to get profitable (which they aren’t) or invest in growth, which is evidence in marketing spend increasing at 16% per quarter.
Now that we have a model for their business it can serve as a tool to measure their business performance over time and push on what factors are at play. What didn’t show up (yet) is the potential market size and impact of competition. As Spotify grows and streaming matures they will find growth harder and will have to prove that they can retain customers.
In Closing
Using systems modeling for businesses is a helpful way to evaluate the performance and which causal loops are at play. I use this to identify metrics to monitor and where to focus investments. While it can be time consuming to build out, the payoff is that you understand your business and how it works. I did a model for product development and it was very helpful to identify how decisions flow through an organization and identify areas for improvement.
I recommend you take time to think about your organization or a product you are bringing to market and what are the dominant feedback loops. Understanding this model can help you to craft a strategy that is dynamic and looks beyond 1st order effects.
References
For this course above Week 2 lays out the principles
