The standard pitch for ML-driven VC sourcing is correct in the aggregate and wrong at the margin where it matters. Hiring velocity, web traffic, GitHub commits, app store ratings, these signals outperform gut intuition across the full universe of startups. The problem is that the full universe is not your portfolio.
Stage-agnostic scoring models are systematically miscalibrated because the signals that predict success at seed have a different shelf life than the signals that predict success at Series A. Using a stale signal isn't neutral. It's noise dressed as insight.
THE DECAY CURVE IS REAL
Hiring velocity is the clearest example. At seed, a team that is aggressively hiring before revenue is a leading indicator of founder conviction, capital efficiency expectations, and organisational ambition. The signal is fresh and predictive, founders who build fast teams early tend to build fast companies.
By Series A, hiring velocity has largely decayed as a signal. The companies that are hiring fast at Series A are often the ones that raised a lot, not the ones that are executing well. Capital causes hiring, not the reverse. The signal that was predictive at seed is now confounded by the round itself.
The same logic applies in the other direction. Revenue retention like net dollar retention, expansion rate, churn cohorts is meaningless at seed because there is no revenue to retain. It's a powerful signal at Series A and a critical signal at growth. Stage-agnostic models that include revenue retention in the seed-stage scoring layer are not ignoring this fact, they are giving it zero weight implicitly, which is different from having thought about it explicitly.
WHAT STAGE-STRATIFIED STACKS LOOK LIKE
The funds winning the sourcing game are not running one scoring model. They're running signal stacks calibrated to stage, with deliberate choices about which inputs are predictive at which moments.
Seed signal stack: Team composition and background data (founders who have built before in the domain), early hiring decisions (which functions they hired first, how fast they hired after funding), product shipping cadence (GitHub activity, product launches, beta user counts), and network graph positioning (who is on the cap table, who is advising, who they have access to).
Series A signal stack: Revenue trajectory and retention shape (not just the number, but whether the curve is accelerating or flattening), sales cycle characteristics (ACV relative to sector, time-to-close relative to cohort), product engagement depth (not DAU, but whether the product is in the critical path of the customer's workflow), and competitive positioning relative to named alternatives.
The transition between these stacks matters as much as the stacks themselves. There is a window between seed and Series A, typically between $500K and $2M ARR where the seed signals are losing predictive power before the Series A signals have enough data to be reliable. This is the hardest scoring environment. The funds that perform here supplement quantitative signals with qualitative inputs they've structured: reference calls that follow a consistent schema, founder conversation recordings analyzed for linguistic patterns, board update language that gets embedded and monitored.
THE PRACTICAL IMPLICATION
If your sourcing model was trained on a corpus that includes seed, A, and B outcomes without stage stratification, it has learned signal weightings that are incoherent. It is applying the predictive logic of seed-stage signals to growth-stage companies and missing the predictive logic of growth-stage signals at seed. The model is not wrong — it is systematically miscalibrated in ways that are hard to see in aggregate and expensive in individual decisions.
The fix is not to discard the model. It is to stratify it. Train separate models for each stage, or apply stage as a conditioning variable that reshapes signal weights rather than treating it as one feature among many.
The funds that have done this will tell you it is a materially different output. The companies that score well on a seed-stage model look different from the companies that score well on a stage-stratified model at the same funding level. That difference is where the alpha is.
THE BOTTOM LINE
Alt-data sourcing works. The evidence is solid enough that this is no longer the argument. The argument is whether you are running a model that understands signal decay or one that treats all inputs as equally predictive regardless of when in the company lifecycle they're measured.
The signal decay curve, the half-life function of each alt-data input as a function of company stage is not a theoretical refinement. It is the difference between a sourcing model that finds the companies that already look like winners and one that finds the companies that will become winners. At seed stage, those are very different lists.
Your signal is fresh at seed. It's stale at Series A. Price them accordingly.