Building a Scoring Algorithm
Combine signal into a defensible numeric score that orders the pipeline.
A scoring algorithm is the simplest way to impose order on a pipeline of thousands. It does not make investment decisions — it ranks. The point is to give the partnership a single number that says look at this one before that one, and a written rationale that says why. Done well, it lets the team argue about weights and rubrics instead of arguing about each company.
The gates that run before any score
Before the algorithm scores anything, run a small set of binary checks that make the question of how good a company is irrelevant. If the company fails a gate, it gets a zero or a flag — not a score.
Outside the jurisdictions you can actually deploy into → out. Binary check against a country list.
Already public, acquired, or out of business → out. The company has moved past the point where you can write a check.
No funding history or investor data at all → flag for manual review, do not fabricate a score.
Start with v1: three criteria you can actually defend
The version of this algorithm that works — the one you can run across thousands of companies and trust — is built on three criteria: thesis fit, investability, and investor quality. Each company is scored 1 to 4 on each criterion. The three sub-scores get weighted, averaged, and rescaled to a number out of ten.
- Thesis fitWeight 50%
What it asks. Does this company sit inside the sectors, technologies, or business models we actually invest in?
Why it works. If the answer is no, nothing else matters. This is the biggest weight because it is the cheapest, most defensible filter.
Signal used: Company description, structured activity summary, web search.
- InvestabilityWeight 30%
What it asks. Can this company absorb the size of check we typically write?
Why it works. A great company at the wrong stage is not actionable. Capital raised, last round size, and last round date triangulate stage reliably.
Signal used: Total raised, last round size, valuation, last round date, round type.
- Investor qualityWeight 20%
What it asks. Have credible institutional investors already done the work of backing this company?
Why it works. Borrowed conviction. Match the cap table against a curated list of investors who actually understand your category.
Signal used: Cap table investors vs. a hand-curated list of top firms in your space.
The math is deliberately simple
What the score means in practice
- 8.5 – 10High priorityStrong on all three. Worth a partner's attention this week.
- 7.0 – 8.4Worth a lookTwo of three are strong. Read the rationale to see which one to test.
- 5.0 – 6.9SituationalReal but bounded. Needs a thesis case or a timing shift.
- < 5.0DeprioritizeWrong stage, wrong sector, or no institutional signal.
Why v1 is enough — and what makes it work
The three v1 criteria all share something important: they are cheap to compute, easy to defend, and hard to game. Thesis fit comes down to a clear sector framework and a short company description. Investability is mostly arithmetic on funding data. Investor quality is a fuzzy match against a list your partnership wrote down. None of them require deep judgment about the company itself — they ask whether the company belongs in the part of the market you actually fish in.
That makes v1 the kind of algorithm you can run across tens of thousands of companies, re-run when new data arrives, and trust to be consistent today, next month, and next year. That consistency is the product. A score that means the same thing across the whole pipeline is more valuable than a score that is sometimes brilliant and sometimes nonsense.
v2 sounds great — and it is genuinely hard
Sooner or later someone will say: shouldn't the score also reflect defensibility, market size, and team? They are right that those things matter. They are usually wrong that the score is the place to capture them.
The three v1 criteria are mostly structural — you can compute them from data the world has already published. The v2 criteria are not structural. They are judgment dressed up as a number, and that judgment is exactly what your partners get paid for. A couple of concrete problems:
Calling a moat from a paragraph of public text is genuinely hard. Most companies look defensible in their own pitch and undifferentiated under a microscope. The model drifts to a comfortable 3 unless you write very strict rubrics — and even then it disagrees with your team half the time.
Sizing a market from public data has more noise than evaluating a cap table. Frontier categories have no clean TAM. The line between a $5B and a $15B market — the line that actually drives whether you would deploy capital — is right where the algorithm is least reliable.
Founder data is wildly uneven. Some people have a decade of public footprint, others nothing. The model penalizes the second group for being invisible, not for being weak. The errors are systematic, not random.
The whole value of a score is that it means the same thing on company #4,832 as it did on company #12. Subjective criteria break that property — the same company can score differently depending on how the web indexed it that week.
None of this means v2 is wrong in principle. It means the cost of getting it right — careful rubrics, regular calibration, honest reporting on where the algorithm is uncertain — is high, and most firms underestimate it. The honest framing: v1 orders the pipeline, v2 is a research project. Ship v1 first. Let the partnership feel what consistent ranking does to triage. Build v2 only once v1 is fully wired into how the team actually works.
Operating principles for any version
- Always return a rationale. A number with no explanation gets ignored. Two or three sentences per criterion is enough to make the score auditable and to let a partner disagree with it productively.
- Renormalize when a criterion is missing. If you do not have the data to score one of the criteria, drop it from the weighted average and rescale the others. Do not punish a company for a gap in your data.
- Make the score re-runnable. Funding rounds close, investors change, descriptions get rewritten. The score for a company today should not be the score forever — re-run on a schedule and on new signal.
- Treat the score as triage, not verdict. A high score earns attention. A low score earns a deferral, not a death sentence. The algorithm is a tool for ordering work, not for replacing it.