Business Concepts for Data Scientists — Finance/Economics (Part I)
Data Science Renee did a popular Q&A on Twitter for aspiring data scientists and one tweet stood out to me in particular:
After doing a little poking around, I couldn’t find any articles that were crash courses in common business concepts for a data science audience. Instead of doing more digging, I decided to write this working draft with a list of concepts I’ve encountered during my career, definitions with sources for depth, and brief examples for why data scientists in particular should know about and care about them. Feel free to skim through to the topics most interesting or relevant to you.
Get Jarus Singh’s stories in your inbox
Join Medium for free to get updates from this writer.
Note that various industries will have additional concepts that are specific to them — so if you’re planning on breaking into those industries, more targeted study will be necessary. This list is likely most useful for data scientist generalists and those working at small companies; folks working in very niche areas in data science will probably find this less useful as the problems they’re working on are very specific and success is usually well defined. Also, this list is by no means exhaustive, please leave a comment or let me know on Twitter if you have suggested additions.
Finance/Economics Concepts for Data Scientists
- Revenue (and sources of): Income generated from normal business operations such as selling goods and services. For some businesses, like lemonade stands, it’s simple to calculate: total lemonades sold x price of lemonade. For other businesses, especially large tech companies, it can be more complicated. Some businesses have multiple ways of generating revenues, a common combination for apps is subscription revenue from users who pay a one time or monthly fee and advertising revenue generating from serving advertisements for users who use the app for free. In these cases, data science is particularly helpful in estimating the impact of various initiatives on individual and total lines of revenue or predicting future revenue streams. Regardless of your role at a given company, you should be aware of the way or ways in which your company generates revenue and what could affect it positively or negatively.
- Cost (Fixed and Variable): Costs are expenses incurred by a business from normal business operations such as selling goods and services. In our lemonade stand example, that would be the cost of whatever we need to make our lemonades (e.g. water, lemons, sugar, ice). Costs can be broken into a lot of different categories, but a meaningful split I think most people should know are the differences been fixed and variable costs. Fixed costs do not change as a company sells different amounts of goods or services. Variable costs do. One needs to know how revenues and costs change under various scenarios. Generally, data scientists should prioritize knowing what variable costs their company has. Different industries have different cost profiles. Social media companies and game apps have low variable costs — it’s practically free to register new users and serve ads to them after the product has been released. Ridesharing companies have to pay their drivers to provide additional rides. Streaming companies like Youtube and Twitch have to pay content creators more when people watch more videos or livestreams.
- Profit: Revenue minus Cost. In my opinion, this is the most important concept for folks to understand. Under neoclassical economic assumptions, the only goal of a firm is profit maximization. Let that sink in for a second. Think about all the things that various folks — data scientists and non-data scientists alike — at your company are trying to accomplish: increasing efficiency, deepening relationships with clients, minimizing service downtime, expanding into new markets, cultivating brand relevance, improving recommendation accuracy, etc. All of these are in service of one goal: maximizing profit. I believe that all employees should understand how their current and potential projects (especially if they have the ability to suggest and prioritize their own work) can maximize profit. One common misstep I’ve seen among data scientists and data science management is to have too narrow a focus on what they think is important. They may choose to spend lots of time and energy eking out small gains in accuracy of predictive models without ever stopping to think about the value of those gains. They also build black box models for systems when causal inference is more useful because it could inform and improve decision making by management. Sometimes, instead of having somebody do that translational work, where data science projects are picked with business outcomes in mind, underperforming data scientists are cut before they’re even told what’s going on. Data scientists (and frankly, everybody) should understand how companies generate (or plan to generate) profit and how their work specifically can affect and maximize profit generation. Note: This is not an endorsement of the neoclassical view of a corporation. I think it’s important to reference for its simplicity and because it’s been the dominant view for a long time. Recently, the Business Roundtable updated its definition of a corporation to be more expansive.
Note on Company Goals: Although I’ve just stated above that the ultimate goal of any business is profit maximization in the long term, depending on your specific role and stage of the company you’re at, your short term priorities may be different. A simplification of company lifecycles is that they go from idea/product->audience growth->revenue growth->profit optimization (in particular, cost management). You should both know what stage your company is in and what role you play in its success. This isn’t to say that data scientists can’t or shouldn’t be working on product improvements (the early stage primary focus) at a late stage company — healthy companies are continuously improving their products — but your success should be relative to profit maximization (a late stage goal) as opposed to say, audience growth. Imagine only presenting user growth numbers from your work at a late stage company to a CEO who is most concerned with how they’re going to hit their quarterly profit target. Not a recipe for career success! If you have some discretion in the sorts of projects you can take on, try to align them with what’s most needed for the company you work for given the stage it’s in. If you’re not sure, I’d recommend asking around. I’ve noticed that lots of technical folks shy away from opportunities to talk to their company leadership, and it can set them apart from their peers if they can do so intelligently and constructively.
- Return on Investment (ROI): (Gain from Investment — Cost of Investment) / Cost of Investment. This formula is applied at an investment by investment basis. As a simple example, if a chain restaurant opens a new location, in the long term, they would want that location to deliver more profit than it cost to open. In some instances in the data science world, it can be just as easy. If we have a data scientist working on recommendation systems at an ecommerce company, they would want the costs of deploying a new model (including the data scientist’s compensation) to be larger than the increased profits that the new model brings. However, many data scientists work on types of projects where this is much more difficult to do. How does one quantify the gains from improved anomaly detection, marketing segmentation, or even an automated dashboard that displays real time metrics? The data scientist (or their manager, or a Product Manager, or leadership) needs to understand how to translate the measurable impacts of these projects (e.g. 20% more key metric anomalies caught) to business impacts (e.g. 5% higher profits). Doing so for each type of data science problem is beyond the purview of this article, but I hope to address it in future ones.
- Unit Economics: This is a style of analysis where profits are examined on a per unit basis. The idea here is that going from 99->100 paying customers may have a different effect on profits than 999->1,000. This can be due to factors such as semi-variable costs or vendor bulk discounts. In summary, don’t assume that adding the same number of additional purchases or amount of ad revenue always has the same effect on profits. You need to treat each situation individually to get accurate answers.
- Counterfactual Analysis (also known as But For Analysis): This is an important method of financial or economic analysis, and data scientists are often doing a version of this themselves without realizing it! A counterfactual analysis asks the question: what would have happened if an intervention hadn’t happened, where an intervention is typically a decision or initiative in the business world. These analyses are used to inform future decisions or initiatives. For example, a company could increase the discount it offers on its products in a handful of locations, and a data scientist or economist could do causal inference with differences in differences to determine whether that increased discount increases profits in the long term — perhaps by increasing purchasing of additional items. If the analysis shows that the discount is profitable in the long term, the company could offer it across all locations. The same is typically done with AB testing, when at the conclusion of the test the optimal treatment is applied to the entire population of customers.
- Opportunity Costs: The loss of potential gain when one particular alternative is chosen over the others. Complex, but important. Imagine a situation in which a company can dedicate resources towards creating and deploying one data science model for a specific problem, A or B. A will make the company $5M in profit annually, and B will make the company $15M. When viewed as a decision between the two, it seems obvious, pick B! The opportunity cost of not doing so is $10M―the difference in profit between the two. Often, the world is not so simple, however. Imagine you proposed model A and your company was unaware of model B at the time of your proposal, so they took your recommendation. Your model starts to make the company $5M more in profit annually. This fact, viewed in isolation, sounds great…but relative to performance from having implemented model B, which we will assume is proposed by a later data scientist hire, it’s worse. You can use counterfactual analysis and opportunity costs to determine that your model was suboptimal, which could inform future decisions (e.g. if there’s a pattern of your model underperforming relative to your peer’s suggestions with respect to incremental profit, you may want to defer to your peer for problems similar to this one). Even though your model suggestions are profitable for the business if implemented, they are not optimal, and the concept of opportunity cost helps us understand that.
