On founder-product fit

I’ve written dozens of blog posts over the years on how Rebel Fund uses data and algorithms to predict tomorrow’s most successful Y Combinator startups, and ultimately achieve an outsized return for our investors.

My last post On Rebel Theorem 4.0 unveils the latest release of our proprietary ML/AI model, which after 5+ years of development and millions of R&D dollars invested, has proven incredibly powerful at predicting future YC unicorns.

I’m often asked by other investors how they can better leverage their own internal data to make better investment decisions. Telling the full story of how we developed Rebel Theorem would take too long, so I’ll focus this post on how we recently added a set of new features into our model around “founder-product fit” because I think it’s a perfect case study on how to brainstorm, test, and incorporate new features into an ML/AI model.

For context, our latest Rebel Theorem 4.0 model already includes 200+ features that we’ve developed, tested, validated, and incorporated over the years, but we’re always looking for new ways further enhance the model’s predictive accuracy. Now that we have advanced AI tools at our fingertips, we can be more creative with this than ever. Here are the steps we followed to incorporate some new founder-product fit features into Rebel Theorem.

Step #1 — Brainstorm Features

Our data science team started with a simple question: Is the “fit” between a founder and their company predictive of startup outcomes?

On its face, it seems that it should be, but given the incredibly high stakes of seed-stage startup investing, we never add new features to our model without thoroughly backtesting them, no matter how intuitively predictive they may seem. So, we broke down this broad question into 3 categories with 8 testable questions across them.

Domain & Industry Expertise:

1) Does the founder have any significant experience in this industry?

2) How relevant is the founder’s prior work experience to the company they are founding?

3) How relevant is the founder’s academic background to the startup?

Customer & Market Insight:

4) How aligned are the geographies in which the founder has previously operated with the startup’s target market locations?

5) How well does the founder understand the current startup’s go-to-market approach (e.g., SaaS, marketplace, DTC)?

6) Has the founder previously worked with or applied the core technology driving their current startup, such as AI, blockchain, or CRISPR?

Leadership & Organizational Readiness:

7) How clearly does the founder’s career path show progression toward larger scope and higher leadership responsibility?

8) How well has the founder demonstrated adaptability in moving between large-corporate environments and scrappy startup settings?

Step #2 — Prompt AI Model

Next we fine-tuned and prompted an AI model (OpenAI’s o3 model in this case) with lots of internal data we’ve collected on every YC company and founder in history, so it could score each founder-product pair on their fit across these 8 questions. Even though we’re testing qualitative features, we ultimately need numerical scores since a machine-learning model like Rebel Theorem only understands numbers in the end.

I won’t share the specific prompts we used because they’re both proprietary and quite technical, but the overall idea was to give the model clear instructions and as much relevant context as possible on the founders and their companies so it could score them accurately. Here’s the scoring scale we used for illustration:

1 — not related at all (e.g. founder worked in sales and product is a developer tool)

2 — vaguely related (e.g. studied CS working and on a B2B vertical SaaS)

3 — very related (e.g. worked as an AI researcher and is starting an AI start-up)

4 — uniquely suited to be working on this startup (e.g. has a PHD or doctorate in biomedical devices and is creating a biomedical device start-up)

Once we got the scores back from the AI model, we could start testing how predictive they are of YC startup outcomes.

Step #3 — Test Scores

Following our internal testing, the team prepared dozens of charts illustrating the usefulness of these new founder-product fit features.

Here are a few examples:

This chart shows the frequency distribution of scores across each of the features that we tested — essentially how common different scores are for each feature. For example, we see it’s quite common for founders to have some background in the same geography of their startup’s target market, but it’s more rare for founders to have demonstrated adaptability moving between large-corporate environments and scrappy startup settings.

This chart shows how predictive each founder-product fit feature is of startup success. The clear winners are education (i.e., how relevant the founder’s academic background to their startup) and leadership progression (how clearly the founder’s career path shows progression toward higher leadership responsibility), at least when the scores are very high or low, as scores in the middle are mostly noise.

Given the steep power law of venture capital returns and how rare great investments are, sometimes it’s less about picking winners than avoiding losers. So, we look at not only how good new features are at predicting startup success, but also at predicting which startups are most likely to fail.

This chart shows how good each founder-product fit feature is at predicting which startups will likely fail — where they really seem to shine! Notice the higher dispersion in the whisker plots vs the previous success chart. A multivariate analysis found our founder-product fit scores are about 10x better at predicting which startups will likely fail than which will succeed.

What this tells us is a lack of founder-product fit is much more detrimental to a startup than an abundance is helpful. Companies that showed poor founder-product fit around geography, go-to market (i.e., how well the founder understands the startup’s go-to-market approach) and work experience (i.e., how relevant the founder’s prior work experience is to the company) were much more likely to die than their peers.

Finally, this chart shows the statistical significance of each feature when it comes to predicting startup failure, which as we’ve discussed, they’re better for than predicting success. In order for a feature to make it into our model, it needs to not only prove predictive of certain startup outcomes, but also with high statistical significance (i.e., beyond mere chance). When it comes to predicting startup failure, all of the founder-product fit features proved statistically significant, though to varying degrees.

Step #4 — Retrain Model

Once we’ve thoroughly tested new features and know them to be both powerful and statistically-significant predictors of YC startup outcomes, we retrain our core Rebel Theorem model incorporating these new features. This way the model can look at the new features alongside the 200+ other features it’s already been trained on to holistically predict YC startup outcomes, and we can see the full impact of the new features on overall model performance. The combined impact of these new founder-product fit scores proved quite dramatic, even though each feature taken in isolation had relatively little predictive power.

So, that’s how our new founder-product fit features made it into Rebel Theorem. When I think about how exhaustive and scientifically-rigorous our process is for adding new features to our model, and how incredibly predictive it is these days of YC startup outcomes, I wonder how I ever made investment decisions without it.