CASE STUDY - VENTURE CAPITAL DEAL SOURCING
Months of custom modeling, matched in a single call.
A venture capital firm's own deal-ranking model took months to build. NEXUS matched it with a single fit-and-predict call - no feature engineering, no tuning - and found more of the deals that went on to win.
The Task
Rank deals by predicted MOIC
Build Time
Hours, not months
Extra Winners Found
+36 over 5 years
Months of tuning to rank incoming deals
The firm scores every inbound and referral deal against its expected 3-year return, so the investment team can spend its time on the roughly 300 to 500 companies a year worth a closer look. The model supports that judgment call - it doesn't make it for them.
To produce that score, the firm's data science team trained a LightGBM deal-ranker on 69 features spanning web traffic, funding and company data, hiring signals, and investor track records - with a custom asymmetric loss, categorical encoding, and results averaged across 5 random seeds to damp fit-to-fit instability. Getting it there took weeks to months, and it kept demanding attention every time the deal flow shifted.
The same raw data, in a single step
NEXUS's pre-trained foundation model was pointed at the same 69 features, exactly as they were - no feature engineering, no categorical encoding, no hyperparameter tuning. One fit-and-predict call, done in hours instead of weeks. Both models were then tested the same way: ranking deals by predicted return and checking those rankings against what the deals actually returned, across five years of real outcomes, 2016 through 2020.
More winning deals, every year measured
Measured Outcomes
+8%
higher returns on the top 100 deals it ranked highest (8.54x vs. 7.91x)
414 vs. 378
deals over 5x returns found in the top 250 - NEXUS vs. the existing model, across 5 years
~$300M
estimated added value on a typical $500M 2020 deal batch, just from that 8% MOIC gain
NEXUS came out ahead every single year of the five-year test, and the gap was widest for the smallest, highest-conviction shortlists - exactly where a stronger call matters most. Same data, same test, no tuning: 36 more winning deals put in front of the team, or roughly 7 extra deals a year returning 5x or more - real capital gains the in-house model's shortlist would have missed.
Precision at the Top 250, Every Backtest Year
Eval year
In-house P@250
NEXUS P@250
Improvement
2016
18.4%
20.8%
+13.0%
2017
22.0%
26.4%
+20.0%
2018
27.6%
29.2%
+5.8%
2019
40.4%
44.4%
+9.9%
2020
41.2%
44.8%
+8.7%
Average
29.9%
33.1%
+11.5%
NEXUS matched or beat the in-house model across 94% of shortlist sizes tested (top 50 to top 500), every year from 2016 to 2020.
Fewer moving parts, more deals surfaced.
NEXUS reached the same bar without the months of setup, freeing the team to spend that time elsewhere. It also means the shortlist keeps working without depending on any one person's pipeline, and it surfaced deals - including two 50x-plus outcomes - that the in-house model ranked far too low to act on. Same data, same test, run head-to-head across five real years - a comparison the team could trust.
