When Rebel Fund began our journey in 2019, our thesis was simple:
Y Combinator startups are the best investment class in venture
YC unicorn founder partners can best select/access the top YC startups
ML (+ later AI) models can best screen startups for partner interviews
Over the past 7 years, all of these have played out. We’ve now seeded over 300 top YC startups, including several unicorns, and established a reputation as one of the most sought-after investors in the YC ecosystem.
We’ve also invested millions of dollars into developing the world’s most sophisticated ML/AI algorithm for predicting YC startup success at the pre-Demo Day stage. I’ve written about our various algorithm iterations over the years, most recently Rebel Theorem 4.0 and our new AI composite scores, and now I’m pleased to announce Rebel Theorem 5.0.
As with its predecessors, RT 5.0 is trained on millions of datapoints on every YC startup, founder, and outcome in history. We track and measure every company and team across hundreds of dimensions, and have identified the exact combination or characteristics (or “features” in ML speak) that most accurately predict YC startup success.
Given how much we’ve invested into developing RT 5.0 and how powerful it’s become, I’m going to be more tight-lipped about the inner workings of the algorithm than in my previous posts. Today’s post will focus more on how we validate and use the algorithm than how it works under the hood.
To start, let me explain what the latest algorithm actually does: predicts a company’s likelihood of enjoying a Series A markup
A lot investors are surprised by that, because Series A rounds aren’t what drive portfolio returns for seed funds like us — it’s multi-billion dollar exits. However, we’ve found in backtesting that it’s impossible to predict at the seed stage exactly which startups will become unicorns, because unicorns are rare (around ~6% of YC startups) and there’s too much noise in the data to identify them pre-Demo Day, even with a model as robust as ours.
However, we can accurately predict a company’s likelihood or enjoying a Series A markup, and then put only the most likely winners in front of our Rebel partners for further diligence. Since ~20% of YC startups that achieve this milestone ultimately become a unicorn, that’s sufficient to enrich our portfolio with enough unicorns to be a top-decile seed fund.
Rebel Theorem 5.0 model performance
Now the billion dollar question: how well does the model work?
Let’s start by comparing it to its predecessor, Rebel Theorem 4.0 + our AI composite scores:

This chart compares the composite prediction scores of RT 5.0 (blue) with RT 4.0 (orange) in terms of how many future winners were predicted per 250 companies interviewed, and how many ‘dead’ companies slipped through. This backtest was based on 2023 & 2024, which is why so many companies haven’t yet been declared winners or dead. As you’ll see, the newer model produced a modest ~7% improvement in predicting winners and a more dramatic ~55% reduction in false-positives.
More importantly though, the ~7% improvement in predicting winners is concentrated heavily on the mega-winners that really drive returns for seed funds like us.

This chart below shows what our actual Rebel partner interview dockets would have looked like those years using the new model. As you’ll see, it heavily enriches our limited partner interview slots with future winners and wastes very little of our partners’ time on companies that die (there are less winners thus far in the newer batches just because they’re still young).

Pharmaceutical companies know a drug is working when a larger dose has a greater effect (dose-response curve). Likewise, we know our model is working when a higher success probability score correlates with actual better company outcomes.

Here we can see that the higher a company’s RT 5.0 score, the more likely they were to actually enjoy a Series A markup. The top 10 companies in each batch had an extremely high hit rate of 42% — nearly 3x the batch average.
Of course this begs the question, why not just interview the top 10 companies as ranked by our model? The answer is that while interviewing more companies decreases our Series A hit rate, it increases our chances of hitting a super-outlier that can drive 100x+ returns for the fund. We’ve determined ~50 interviews per batch to be the optimal balance between risk (investing in too many losers) with reward (investing in enough mega-winners). See my post On hitting the bullseye to understand why a diversified portfolio is so important for seed investing.
Here’s the actual gain curve:

Last but not least, I’ll share a chart that shows the correlation between RT 5.0 markup probability scores vs actual Series A markups.

As you’ll see, the model is extremely well-calibrated to its stated objective. When a company gets a markup probability score of ~30% then their chances of enjoying a Series A markup really is ~30%.
I could share more charts and graphs, but the bottom line is our model works. You’d be hard-pressed to find another ML/AI model for predicting startup success that’s nearly as accurate and well-validated.
Ironically, the model doesn’t have to work as well as it does for us to maintain world-class fund performance — just not under-performing YC averages is enough. YC now has $1.3T in portfolio company value, well over 100 unicorns, a 3x unicorn rate vs non-YC, and the 13x TVPI (2014–2018 vintages) for Demo Day investors overall.
I know many of you are curious how Rebel Theorem 5.0 actually works, but unfortunately I can’t share anything beyond what’s in my previous posts about what actually predicts YC startup performance. In today’s environment, the stakes are just too high :-)

