Building the Cost-Benefit Case for a Data Purchase

A researcher who has confirmed a dataset is structurally sound and shows a genuine, statistically credible signal is only halfway to a purchase recommendation. The remaining question is not "does this work" but "is this worth what it costs" — and that question has to be answered in dollars, on both sides, or the decision gets made on vibes. A data budget committee that hears "we found real alpha" approves a purchase far less reliably than one that hears "this adds an estimated $310,000 a year in net PnL against a $120,000 annual cost."

What goes on each side of the ledger

Cost is usually more than the sticker price. Annual licence fees are the visible number, but a full cost estimate includes onboarding engineering time (entity resolution, pipeline building), ongoing storage and compute, and the ongoing maintenance burden of a feed that needs monitoring for schema changes and outages. A dataset that looks cheap at $50,000/year can cost far more in the first year once a few weeks of a researcher's and an engineer's time are priced in.

Benefit is the marginal PnL the dataset's signal is expected to add to the book, net of trading costs, and — critically — marginal to what is already owned. If the new dataset's signal is 70% redundant with an existing factor, only the incremental 30% belongs on the benefit side; crediting the whole gross signal double-counts value the book already had.

$\text{net value} = \left( IC_{\text{marginal}} \cdot \sigma \cdot \text{AUM} \cdot \text{breadth factor} \right) - \text{cost}_{\text{total}}$

In words: the marginal benefit scales with how much new predictive power the signal adds (its marginal, not gross, information coefficient), the volatility and size of the book it's applied to, and a breadth adjustment for how often the signal actually gets to make a call. Total cost is licence plus onboarding plus maintenance. The purchase is worth making only if net value is positive by a comfortable margin — a small positive estimate isn't worth the operational commitment, given how uncertain the benefit estimate typically is.

A worked example

A sentiment dataset costs $180,000/year in licence fees, with an estimated $70,000 in one-time onboarding engineering and $15,000/year in ongoing maintenance — total year-one cost around $265,000, and $195,000/year thereafter.

Research shows the raw signal has a monthly IC of 0.035. Orthogonalised against the existing signal library (which already includes a news-sentiment factor with meaningful overlap), the marginal IC drops to 0.014 — most of the raw signal was redundant. Applying that marginal IC through the desk's standard alpha-to-PnL translation, at the strategy's current AUM and turnover budget, yields an estimated marginal PnL contribution of roughly $310,000/year before costs.

	Year 1	Year 2+
Licence fee	$180,000	$180,000
Onboarding (one-time)	$70,000	—
Maintenance	$15,000	$15,000
Total cost	$265,000	$195,000
Estimated marginal benefit	$310,000	$310,000
Net	+$45,000	+$115,000

The case is positive but thin in year one, and considerably more attractive from year two onward once onboarding is a sunk cost. This framing — separating the one-time cost from the recurring one — is what lets a committee correctly evaluate a purchase that looks marginal in aggregate but is actually a good multi-year bet.