What 172 Deal Memos Taught Me About Pattern Recognition

When I was building NUVC's evaluation engine, I needed training data that most people don't have access to: real VC deal memos from investments with known outcomes — companies that had since IPO'd, been acquired, or failed.

The process of sourcing, anonymising, and analysing 172 of these documents took months. The findings took longer to sit with — because several of them contradicted assumptions I'd built my own thinking around.

I want to share what the data showed. Not as a finished research paper (that targeting publication separately), but as a practitioner's account of what surprised me, what confirmed existing intuitions, and what I think it means for how investors should allocate their attention.

What the Research Was and Wasn't

172 deal memos is a small sample for statistical confidence on anything. I want to be clear about that.

The memos came from a range of fund vintages, geographies, and sectors — with a deliberate oversample of Asia-Pacific and emerging market deals, which reflects NUVC's focus. They included both well-known outcomes (Canva, Stripe, SpaceX validated the model; Theranos and FTX stress-tested it) and less-known companies where the outcome was unambiguous but the journey wasn't public.

This isn't peer-reviewed research. It's a practitioner's dataset. But it's the most systematic analysis I've been able to do on questions that are usually answered by gut feel and attribution bias, and I think the patterns are worth discussing.

Finding One: Product Execution Quality Is the Most Predictive Signal

Of all the factors I tried to operationalise, product execution quality had the strongest correlation with positive outcomes — r²=0.77 across the dataset.

By "product execution quality" I mean: does the product, as it exists at the time of the investment memo, actually do what the company claims it does? Is the quality of execution evident — in the UX, in the technical architecture, in early user behaviour — or is the company primarily selling a vision of what the product will be?

This sounds obvious stated plainly. It wasn't obvious in the memos themselves. The memos that preceded failed investments were full of credible-sounding product descriptions that, in retrospect, didn't reflect the reality of the product as it existed.

The pattern I kept seeing in the failure cases: the memo described the vision compellingly, and the due diligence didn't verify the gap between vision and execution. The investors relied on the founders' description of the product rather than using the product themselves, talking to early users, or having someone technically qualified assess the architecture.

Finding Two: Team Composition Is Far Less Predictive Than Claimed

Team composition — the pedigree, track record, domain expertise of the founding team — was the most frequently discussed factor in the deal memos. It was also the least predictive of outcomes, with an r² of 0.31.

This is a significant gap, and I spent a long time stress-testing it because it runs counter to so much conventional VC wisdom.

What I think explains it: team composition is legible and defensible. You can source it, verify it, and explain it to an LP in one sentence. "Top team, top school, prior exit" is a sentence that closes investment committees. "I believe this founder has the kind of conviction that will carry them through the hard moments" is much harder to defend.

So investors over-index on what they can write down and defend, and under-index on what requires judgment they can't easily articulate.

High-pedigree teams also have a particular failure mode: they're better than average at pitching, at managing investor relationships, and at buying time when things aren't working. These skills delay the signal that something is wrong — which can make the eventual failure more expensive.

This doesn't mean team doesn't matter. It clearly does. But the specific markers investors use as proxies for good team quality (academic credentials, prior employer prestige, warm intros from famous investors) are weaker predictors than the industry assumes.

Finding Three: Conviction Is the Most Underweighted Signal

I coded "conviction" in the memos not as a single variable but as a cluster: evidence that the founder was building this regardless of funding, that their engagement with the problem predated the company, that they'd made personal sacrifices specific to this work.

Conviction showed up in the language of the memos — you can feel it in how the investor describes the founder — but it almost never appeared as a formal evaluation criterion. It wasn't weighted. It wasn't scored. It was a colour in the prose.

When I correlated conviction-cluster signals with outcomes, the relationship was clear. Companies with strong conviction signals in their initial memo substantially outperformed, particularly in cases where early product-market fit wasn't obvious.

The mechanism makes sense: conviction is what carries founders through the periods when the product isn't working yet, when the metrics don't support the narrative, when the early adopters aren't enough to show institutional momentum. In those moments, the credential-heavy founder and the conviction-driven founder diverge. One reaches for a pivot. The other goes back to the user and tries to understand.

Finding Four: Market Timing Is Discussed But Not Operationalised

Almost every memo had a section on market timing — why now. In practice, most of these sections were rationalised after the investment thesis was formed, not before.

The ones that were actually useful for predicting outcomes shared a characteristic: they identified a specific, recent enabling change — a regulatory shift, a technology unlock, a cost inflection — and articulated precisely how that change made this company possible now in a way it wouldn't have been two years ago.

The generic "markets are ready" timing arguments — which referenced general trends (digital transformation, AI, climate awareness) without specifying the precise enabling mechanism — correlated weakly with outcomes.

Being early on timing is indistinguishable from being wrong in the first two years. The memos that correctly identified precise enabling mechanisms gave the investors a clearer framework for knowing which signals to watch for and when to get nervous.

Finding Five: Prestigious Backing Is a Weak Positive Signal That Decays

Having a tier-1 fund as a prior investor is a positive signal at the pre-seed and seed stage, where it can indicate that smart people have looked at this and believe in it.

By Series A and beyond, it becomes a noise signal and occasionally a negative one. The late-stage failures in the dataset (companies that raised significant capital before failing) were over-represented among companies with prestigious early backers — partly because prestige eases subsequent raises, which can delay the feedback loop that would otherwise force a pivot or a wind-down.

This is not a finding I would have predicted before the analysis.

What I Changed in How I Evaluate Companies

I use the product before reading the deck. Not to grade it as a product, but to see the gap between what it is and what the memo will claim it is. That gap is the most useful thing I learn in the first hour of any new evaluation.

I ask founders to walk me through the hardest thing they've built, not the best thing. The hardest thing tells me what they're capable of when the answer isn't obvious.

I pay less attention to the team slide and more attention to how the founder answers the question: "What have you tried that didn't work?"

The answer to that question — its specificity, its honesty, the speed with which they identify real failures versus performed failures — tells me more about their capacity to navigate the company to an outcome than anything on their CV.

The Limit of Pattern Recognition

Pattern recognition trained on historical data will always underweight genuinely novel opportunities — by definition. The company doing something that has never worked before won't match the patterns of companies that have.

The job of an investor isn't to run pattern recognition at scale. It's to use pattern recognition for triage, and then apply genuine judgment in the cases where the patterns are most useful and most likely to mislead.

The 172 memos taught me which patterns are reliable and which ones investors are using as substitutes for harder thinking. That distinction — between reliable signal and defensible proxy — is, I think, the most useful thing I got from the exercise.

---

Related: The Scoring Model Trap — on why quantifying these signals makes them less useful, not more. Reading Founder Conviction — the deep dive on the finding that surprised me most.

Tick Jiang is the technical co-founder of NUVC (nuvc.ai), an AI-native venture capital intelligence platform built in Melbourne. Her empirical research on VC decision-making is targeting publication at ICAIF, FAccT, and AAAI. She writes on capital, AI, and building across the Asia-Pacific.