What if startup success wasn’t just about gut feelings, luck, or storytelling — but data? And what if machines could now identify which startups will succeed, before any revenue, media buzz, or product launch?
A groundbreaking academic study titled “To shine or not to shine: Startup success prediction by exploiting technological and venture-capital-related features” (2025) explores exactly this. Conducted by researchers at several institutions and published in Information & Management, the study uses machine learning to forecast which early-stage startups are most likely to succeed — and it does so with striking accuracy.
Machines Can Read the Signals Founders Don’t See
The researchers built predictive models using publicly available startup data — GitHub repositories, programming languages used, software development metrics, investor details, funding rounds, and more. They tested over 2,000 startups and trained algorithms to distinguish between those that eventually exited successfully (through acquisition or IPO) and those that did not.
The result? Their model achieved an Area Under Curve (AUC) of 0.85, meaning it was highly accurate in identifying potential winners. The most powerful predictors were:
- Technological signals: frequency and quality of code commits, diversity of tech stacks, use of modern frameworks.
- VC-related features: the reputation and network of early investors, size and timing of funding rounds, and the startup’s position within the venture ecosystem.
In short, success leaves a trail — and algorithms can read it.
Why This Matters for Venture Capital
Traditionally, venture capital decisions have leaned heavily on soft signals: the charisma of the founder, personal networks, and the elusive “gut feeling.” But as funding becomes more competitive and the pool of startups grows globally, VCs face the risk of overlooking high-potential companies — especially those outside of major hubs like Silicon Valley.
Algorithms can benchmark startups against thousands of peers—for example, comparing GitHub activity levels or analyzing investor composition—giving VCs scalable insight into a startup’s promise.
Such models also help mitigate cognitive biases by focusing on quantifiable traits rather than familiarity or media hype.
Ultimately, they streamline the process by automating the initial screening phase, allowing human investors to concentrate on in-depth evaluation of the most promising candidates.
Some VC firms already use internal scoring systems, but academic research like this provides peer-reviewed evidence that it works — and could soon become industry standard.
The Risks of a Data-Driven Future
Of course, no model is perfect. The researchers themselves acknowledge that algorithms trained on historical data might reinforce existing biases — favoring startups with connections to elite investors or certain geographies.
There’s also a danger of startups gaming the system: open-sourcing polished code, inflating GitHub activity, or mimicking the tech stacks of successful peers.
Still, the authors argue that machine learning isn’t meant to replace human intuition — it’s meant to augment it. When used as a complementary tool, it can surface hidden gems that the human eye might miss.
A New Era for Startups and Investors
We often romanticize the idea of a brilliant founder building in a garage. But what if the real breakthrough isn’t in the garage — it’s in the data trail they leave behind?
In a world where success can be predicted, and failure anticipated, the rules of venture capital may be on the brink of transformation. For founders, the message is clear: what you build, how you build it, and who you build it with now shapes your destiny more than ever — not just in the market, but in the eyes of an algorithm silently watching from the wings.
For founders, the message is clear: what you build, how you build it, and who you build it with now shapes your destiny more than ever — not just in the market, but in the eyes of an algorithm silently watching from the wings. And for VCs? The best deal of your life might already be in the dataset — you just haven’t seen it yet.
Prepared by Navruzakhon Burieva
