Deal economics
Contingent data deal value analysis belongs on a probability tree
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Contingent data deal value analysis starts from the payment schedule, not the headline. Draw each event a payment waits on, from commitment and acceptance to milestones and refund windows, as a branch with a conditional probability backed by dated records, then multiply along each path. Test shared causes too: one buyer or one rights defect can sink several deals at once, leaving the expected value unchanged but the downside far worse.
Value drivers
Contingent data deal value
Key takeaways
- A signed contingent deal is worth the probability-weighted payments it can still release and keep, not its headline.
- Multiply stage odds only when each one is conditional on the earlier stages passing; rates borrowed from other deals break the chain rule.
- In the illustrative tree, a 100-unit headline is worth 56.8 units, and the one-probability shortcut overstates it by 23.2.
- Early branches, commitment and acceptance, move value most when the money sits in early tranches; build the register before the sensitivity.
- One shared shock across five deals leaves expected value unchanged but lifts the chance that none pays from 0.6 to 15.1 percent.
- Ask any data licensing company, SourceX included, for dated stage counts with denominators and a list of shared causes.
What is a signed but contingent data deal worth?
A signed contingent data deal is worth the probability-weighted sum of the payments it can still release and keep, not its headline. Each payment waits on later events: the buyer's commitment, an accepted delivery, a closing, a milestone and a refund window that can hand money back. Contingent data deal value analysis prices those events one at a time, in the order the contract sets, and shows the committee exactly what it is betting on.
The investment committee's decision is narrow: how much of a signed but unpaid deal belongs in the underwriting case, and which evidence would move that number. A useful answer needs three things a headline lacks. It needs the payment schedule and its conditions, a conditional probability for each condition with the dated records behind it, and a view of which conditions can fail for the same reason.
Accounting already refuses to treat uncertain consideration as certain. Under ASC 606 (FASB ASU 2014-09), a seller estimates variable consideration with an expected value or a most likely amount, whichever better predicts what it will be entitled to, and includes it only to the extent a significant reversal is not expected. An investor's tree is not a revenue calculation, but it borrows the same discipline: no uncertain payment enters the number at face value.
Acquirers do the same when part of a price depends on later events. Bio-Rad's Form 10-K for fiscal 2025 describes valuing contingent consideration from its Stilla acquisition with a probability-weighted model whose main inputs are the probability of achieving the milestones and the discount rate, reported as a Level 3 measurement because those inputs cannot be observed in a market. The useful lesson for a data deal is the structure, not the inputs.
Which events does each payment wait on?
Each payment waits on a chain of events, and the contract sets that chain, not the pitch deck. Read the agreement for every condition that stands between signature and cash the company keeps, and write each one down as a branch with two outcomes. The usual chain for a data licensing deal has five links:
- Commitment: a signature can open a process without committing the counterparty to buy. Check what the buyer can still walk away from, and until when.
- Delivery and acceptance: a payment tied to acceptance waits on a test. ASU 2014-09 treats a customer's acceptance as a possible sign that control of the delivered asset has passed, which is why an open acceptance clause should never be read as a formality.
- Closing conditions: anything else that must be true before money moves, such as a rights confirmation, a signed amendment or a completed de-identification review.
- Milestones: later deliveries, refreshes or volume thresholds that release further tranches, each with its own acceptance.
- Refund and clawback windows: periods in which money already received can be reclaimed if a rights, privacy or quality defect surfaces after acceptance.
Why stage odds must be conditional, and where borrowed rates mislead
Multiplying stage probabilities is correct only when each factor is conditional on every earlier stage having passed. The chain rule writes a path as the chance of A, times the chance of B given A, times the chance of C given A and B; that identity holds whatever the dependence between stages, but only if each factor is measured on the deals that actually reached that stage.
Borrowed rates break the identity. An acceptance rate drawn from all deliveries, a refund rate taken from another product line or an industry closing rate each describes a different population from the deals that survive this contract's earlier gates. NASA's procedural requirements for probabilistic risk assessment describe the same structure as an event tree: an initiating event followed by success or failure at each pivotal event until an end state is reached. A contingent deal is that tree with a payment attached to each end state.
Drug development shows how much the counting method alone can move an estimate. Wong, Siah and Lo (Biostatistics, 2019) analyzed 406,038 trial entries covering more than 21,143 compounds from 2000 to 2015, following each drug's actual development path rather than estimating from sampled phase transitions, and reported an overall probability of success of 13.8 percent, above the earlier estimate the paper compares it with. Data deals have no published base rates of that kind, so a committee should expect each branch to come from the company's own dated records, path by path, or to be shown as an assumption.
The branch register: one illustrative deal priced path by path
The register below turns one illustrative contract into a tree with a headline of 100 units. Every probability and payment share is invented to show the mechanics; none describes SourceX or any real deal. Stage A releases nothing, stage B releases 60 units, a failed refund window returns those 60 units and ends the deal, and stage D releases the last 40.
Enumerating the five end states gives these illustrative probabilities: the buyer never commits (20.0 percent), the delivery is not accepted (12.0 percent), the first payment arrives and is then refunded (3.4 percent), the 60 units are kept but the milestone fails (19.4 percent), and every stage passes (45.2 percent). The expected value is 56.8 units of the 100-unit headline, and the chance of keeping nothing is 35.4 percent. The common shortcut of multiplying the whole headline by the chance of commitment gives 80.0 units, overstating the tree by 23.2.
The evidence column matters more than the numbers. A probability with no dated denominator behind it is an assumption, and an honest register shows it as a range rather than a decimal. Each row gives the stage, the payment it releases, the illustrative conditional probability, the evidence that would justify a real number and the shared cause to test:
| Stage (given all earlier stages passed) | Payment released (illustrative) | Conditional probability (illustrative) | Evidence that would support a real number | Shared cause to test |
|---|---|---|---|---|
| A. Buyer commits after signature | 0 units | 80% | Signed deals in a stated period that reached commitment, over all deals signed in that period | One buyer behind several deals; a change in that buyer's budget or data policy |
| B. Delivery accepted and deal closes | 60 units | 85% | Acceptance outcomes for deliveries of the same data type, with rejection reason codes | A rights or de-identification defect in the source records |
| C. Refund window passes without a claim | Keeps the 60 units; a failure refunds them | 95% | Claims on accepted deliveries, counted from acceptance to the end of the window | The same rights or privacy defect, found after acceptance |
| D. Milestone delivery accepted | 40 units | 70% | Second-delivery acceptance on deals whose first delivery passed | Supplier capacity; a change in the buyer's specification |
| Result of the tree | Expected value: 56.8 of 100 units | Keeps nothing: 35.4% | Exact enumeration of the five end states, described in the methods note | Correlation across deals that share a buyer or a rights theory |
Which branch should diligence test first?
Test the early branches first: in this illustrative tree, the buyer's commitment and the delivery acceptance each move the expected value more than twice as far as the refund window or the milestone. Moving each input alone across the range in the table, with the other three held at base, shows the order; the refund range is narrower only because a probability cannot exceed 100 percent.
That ranking is not a law of data deals; it falls out of this payment schedule. Where most of the money sits in late milestones, the late branches dominate instead, which is why the register has to be built from the contract before anyone runs a sensitivity.
The refund branch looks small here only because its illustrative failure rate is low. It is still the branch most likely to share a cause with acceptance, since a defect that passes the acceptance test and surfaces later turns a kept payment into a liability. The one-at-a-time ranges are:
| Input moved alone | Low case | High case | Expected value, low to high (units of 100) |
|---|---|---|---|
| A. Buyer commitment | 70% | 90% | 49.7 to 64.0 |
| B. Acceptance and closing | 75% | 95% | 50.2 to 63.5 |
| C. No refund claim | 90% | 100% | 53.9 to 59.8 |
| D. Milestone acceptance | 60% | 80% | 54.3 to 59.4 |
Does a pipeline of contingent deals spread the risk?
A pipeline spreads risk only when its deals fail for unrelated reasons. Five deals built on the illustrative tree have the same expected value whether they are independent or tied together by one shared shock, but their downside is very different, and an expected value cannot show the difference.
Call a deal paying if it keeps at least its first payment, which the tree puts at 64.6 percent. If the five deals are independent, the chance that none of them pays is 0.6 percent. Now add one shared shock, such as the same buyer pulling back from every deal at once, with a 15 percent chance of happening, and raise each deal's own commitment odds so its stand-alone probabilities stay exactly as before. The chance that none of the five pays rises to 15.1 percent, while the expected number of paying deals stays at 3.23 and the pipeline's expected value stays at 284.2 of 500 units.
Reliability engineers met this problem in redundant equipment long before investors met it in pipelines. NASA's PRA procedures guide devotes a chapter to dependent events and common-cause failures, the engineer's name for one cause knocking out components that were supposed to back each other up. In a data business the usual common causes are one buyer behind several deals, one legal theory behind several rights reviews, and one supplier or de-identification route behind several deliveries.
Privacy belongs on that list. NIST's generative AI profile (NIST AI 600-1, 2024) names data privacy as a risk, covering leakage, unauthorized use or disclosure and de-anonymization of personal or sensitive data, and a de-identification method reused across deals can fail across deals. The pipeline outcomes under each assumption are:
| Five illustrative deals | None of five pays | At most one pays | All five pay |
|---|---|---|---|
| Independent deals | 0.6% | 5.6% | 11.3% |
| Shared shock, 5% chance | 5.3% | 8.7% | 13.8% |
| Shared shock, 10% chance | 10.2% | 12.2% | 17.1% |
| Shared shock, 15% chance | 15.1% | 16.1% | 21.6% |
What could make a probability tree mislead the committee?
A tree misleads when it looks more certain than its evidence, and the ways it does so are predictable. Each failure below is a reason to read the evidence column before the expected value:
At a glance
What could make a probability tree mislead the committee
- False precision: confident decimals with no denominators turn a forecast into something that looks like a measurement. Show ranges where the records are thin.
- Missing branches: a renegotiated scope, a partial acceptance or a price change after delivery leaves probability unaccounted for if the tree has no branch for it.
- Rates from the wrong population: an acceptance or refund rate measured on other deals is not a conditional probability for this one.
- Shared causes left out: independence understates pipeline downside, as the five-deal table shows, even while the expected value stays put.
- Timing ignored: the illustrative tree leaves out when each payment arrives. A full model discounts each end state; Bio-Rad's disclosure lists the discount rate as a key input beside the probabilities.
- Demand, price and rights risk: a buyer's appetite for licensed business records, and the price it will pay, can shift with model strategy, cheaper synthetic data or court rulings on training data. Those belong in the commitment branch, not in a footnote.
- Schedule engineering: a seller can split milestones so that early branches look surer than the money behind them. Read the payment schedule itself, not a summary of it.
Questions to put to any data licensing company, SourceX included
Applied to SourceX, the tree is a diligence method, not a valuation. SourceX's deal values, stage probabilities, payment amounts and timing are not public, and this article states none of them.
What is public is the order of some events. No records reach SourceX until the supplying company has agreed the dataset's scope and terms and put its signature on them, and each deal's signed agreement fixes its exclusivity and term. Those facts tell an investor where branches sit; they say nothing about how likely each branch is.
SourceX is raising from accredited investors under Rule 506(c), and its investor page (sourcex.si/investors) is where it shares company and commercial metrics directly. Whichever data company is across the table, these are the questions that fill a register with evidence rather than hope:
- For deals signed in a stated period, how many reached each stage, and what is the denominator at every step?
- Which conditions sit between signature and kept cash in the standard agreement, and which can a counterparty leave without penalty?
- What share of signed value depends on acceptance, later milestones or refund windows, rather than on signature alone?
- Which deals share a buyer, a rights theory, a supplier or a de-identification route?
- How many accepted deliveries later produced a claim, and how long after acceptance did each claim arrive?
- Which probabilities in the plan come from dated records, and which are management's assumptions?
Methods note: how the illustrative tree and pipeline were computed
One script produces the tree, the sensitivity ranges and the five-deal pipeline by enumerating every end state with exact rational arithmetic, so there is no sampling and no random seed. Inputs: a 100-unit headline; conditional probabilities of 80, 85, 95 and 70 percent for stages A to D; 60 units released at acceptance and 40 at the milestone; a failed refund window returns the 60 units and ends the deal. Values are rounded once, half up, to one decimal place.
The shared-shock pipeline makes stage A fail for all five deals with probability s and otherwise raises stage A's probability to 80 percent divided by one minus s, which leaves each deal's own probabilities unchanged; the construction is valid for s up to 20 percent. A deal counts as paying when it keeps at least its first payment. The sensitivity moves one input at a time across the stated low and high cases.
What was not measured: no real deal, industry base rate or SourceX figure enters the model, and timing and discounting are excluded. The cited sources (FASB ASU 2014-09, Bio-Rad's fiscal 2025 Form 10-K, Wong, Siah and Lo 2019, NASA's risk assessment guidance and NIST AI 600-1) support the method, not the inputs. Sources were checked on October 27, 2026; several were verified through search excerpts, and a reviewer should open those pages before publication.
Frequently asked questions
What is contingent deal value in a data licensing deal?
It is the value of payments that will only be made, or only kept, if later events happen: the buyer commits, a delivery passes acceptance, a milestone is met and no refund is claimed. It equals the sum of each possible payment times the probability of the path that leads to it, less anything that could be returned.
Why not multiply the headline by the chance that the deal closes?
Because later payments need more events to go right than the first one. A single probability treats a final milestone as if it were as likely as the first tranche. In this article's illustrative tree, that shortcut overstates value by 23.2 units on a 100-unit headline.
Where should branch probabilities come from when no industry base rates exist?
From the company's own dated records, counted path by path: of the deals signed in a period, how many reached commitment, how many of those passed acceptance, and so on. Where records are thin, show the branch as a range and label it an assumption.
Does a pipeline of several contingent deals reduce risk?
Only if the deals can fail for unrelated reasons. Deals that share a buyer, a rights theory or a de-identification route can fail together, which leaves the expected value unchanged but makes a total loss far more likely than independent odds suggest.
Is this how accountants recognize variable consideration?
It is related but not the same. ASC 606 asks a seller to estimate uncertain consideration with an expected value or a most likely amount and to include it only where a significant reversal is not expected. An investor's tree values the deal for an underwriting decision and is not a revenue figure.
Sources
- FASB's ASU 2014-09 (ASC 606) lets a seller estimate variable consideration by expected value or most likely amount, whichever better predicts entitlement, includes it only where a significant reversal is not expected, and treats customer acceptance as a possible indicator that control has passed. Source
- Bio-Rad's fiscal 2025 Form 10-K describes valuing Stilla contingent consideration with a probability-weighted model whose key inputs are the probability of achievement and the discount rate, classified as Level 3. Source
- Wong, Siah and Lo (Biostatistics, 2019) estimated clinical trial success from 406,038 entries for more than 21,143 compounds (2000-2015) by following development paths, reporting an overall probability of success of 13.8 percent. Source
- NASA's probabilistic risk assessment procedural requirements describe an event tree as an initiating event followed by successes or failures of pivotal events until end states are reached. Source
- NASA's Probabilistic Risk Assessment Procedures Guide (NASA/SP-2011-3421) has a chapter on dependent events and modeling common-cause failures. Source
- NIST AI 600-1 (2024) names data privacy among generative AI risks, covering leakage, unauthorized use or disclosure and de-anonymization of personal or sensitive data. Source
- SourceX's investor page says it shares company and commercial metrics with investors directly and takes investor materials requests. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.