PORTFOLIORAY
Not another dashboard. A predictive model trained on 3,027 quarterly snapshots from 107 public companies—learning patterns that precede success, struggle, and failure. Upload your CSV, get predictions in 30 seconds.
Skeptical? Good. 95% accuracy sounds high because we use time-based validation: train on past snapshots, test on future ones. This mimics real-world use.
Calculate ratios in Excel, compare to benchmarks, make gut calls on which companies need attention.
Learns patterns from 3,027 historical snapshots that preceded outcomes. Spots combinations humans miss.
Example: A company with good margins but specific growth + burn rate combination that preceded 87% of failures in training data. You'd miss it; the model catches it.
These are real ML predictions on a sample portfolio—not mockups.
Initializing ML model...
Not just calculations—pattern recognition across 3,027 data points
Identifies combinations of metrics that preceded specific outcomes in training data.
Human equivalent: Having seen 3,027 company trajectories and remembering which combinations led where.
Detects non-obvious patterns that humans miss—like good metrics masking underlying problems.
Example: Great capital efficiency + specific growth trajectory = 73% preceded sudden declines in our data.
Shows which companies in the training set had similar patterns—and what happened to them.
Why it matters: Context for the prediction—not just a number, but a reference class.
Upload your CSV. Get predictions in 30 seconds.
Predictions are informational signals, not financial advice. Use alongside your own judgment.
Transparent about our data strategy
Anonymized features only—margins, growth rates, outcomes. Never company names or deal terms.
Same pipeline, more data. Takes 5-10 minutes. Accuracy typically improves 2-5%.
New patterns learned help all users. Network effect = competitive moat.
We're looking for ONE VC firm to partner with deeply. You get significant equity and help shape the product. Together, we build the portfolio intelligence platform that becomes industry standard.
Anonymized patterns—not company names. We learn what predicts outcomes.
When you recommend us to other VCs, you're building the value of your equity.
You tell us what VCs actually need. We build it.
Real ownership we determine together. As we grow, you grow.
You're using AI portfolio intelligence before your competitors.
Every VC you bring in improves accuracy for everyone—especially you.
We succeed together or not at all.
Features, not identities.
Can't leak what we don't keep.
You control what's shared. Delete anytime.
Your success = our success.
Looking for a VC firm with strong network and willingness to champion new tools.
Let's Talk →amaan@portfolioray.com
Have questions about data privacy, how the model works, or partnership terms?
Real predictions on a sample portfolio. Click any company to see the AI's full analysis.
Loading ML Model...
This takes a few seconds on first load
✓ Connecting to server
○ Loading ensemble model (3 algorithms)
○ Running predictions on 8 companies
○ Calculating risk scores
Feature importance learned from training data (click for details)
Download sample CSVs based on public portfolio data:
Upload your actual portfolio to see predictions for your companies.
Full transparency on data sources, methodology, accuracy, and limitations.
Running validation through ensemble model...
Testing accuracy on 107 companies (3,027 snapshots)
The test: We show the model a company it has never seen before and ask: "Will this company succeed, struggle, or fail?" It gets it right 95.2% of the time.
Why that matters: If you randomly guessed, you'd be right 33% of the time (1 in 3). Our model is 2.9x better than random guessing.
How we know it's not cheating: We use time-based validation — the model is trained on past data (2015-2020) and tested on future data (2021-2023). It literally predicts the future, just like you would use it.
The ensemble advantage: We don't trust just one algorithm. Three different models vote, and we go with the consensus. This catches blind spots any single model might have.
Our model is trained on 107 public companies with 3,027 quarterly snapshots from SEC EDGAR. Every data point is from official SEC filings—legally required to be accurate.
Includes both headline companies and lesser-known startups with documented outcomes.
✓ 100% Verifiable: Every outcome can be independently verified through official sources. Request citations for any company.
Memorizing "WeWork = failure" then asking "Will WeWork fail?"
That's circular — the model just looks up the answer.
Learn patterns from features, then test on companies never seen during training.
Cross-validation ensures no data leakage.
We extract numerical features that existed before outcomes were known. The model learns: "Companies with [these patterns] tend to fail" — not "WeWork = failure." When you upload your portfolio, we apply the same pattern matching to your companies.
We extract 12 proprietary signals from portfolio data — signals that were observable BEFORE outcomes were known. The model learns which patterns predict success vs failure.
Performance metrics relative to sector benchmarks
Trajectory and momentum indicators
Team, execution, and risk factors
Important: The company NAME is only used for display. The actual prediction uses numerical features only. The model doesn't "look up" if a company is in its training data.
Want the full technical breakdown? Our detailed methodology — including specific features, model architecture, and validation approach — is available during partner presentations. Request access →
Privacy: Your data is processed in-memory and never stored. We don't add your companies to our training data. Each analysis is independent and ephemeral.
Out of 100 predictions, ~95 are right
What you'd get flipping a 3-sided coin
When we show the model a company's data and ask "Will this company succeed, struggle, or fail?" — it gets the answer right 95 times out of 100. Random guessing would only get 33 right.
Key insight: The model is especially good at catching failures (97%). This is what matters most — early warning before a company crashes.
P = Precision, R = Recall, F1 = Harmonic mean of P and R
Four algorithms combined, each catching patterns the others miss. Why ensemble? Different algorithms have different blind spots — combining them improves robustness.
Bottom line: Use these predictions as ONE input among many — not the sole basis for decisions. Combine with your own due diligence, board reports, and founder conversations.
Time-based validation: train on past snapshots (2015-2020), test on future ones (2021-2023). 7 test folds, all between 86.5%-98.9% accuracy. Model is always tested on companies it wasn't trained on.
1000 bootstrap samples provide 95% CI on accuracy. This tells you the range of accuracy you'd expect if we re-ran the experiment.
Isotonic regression calibrates probability outputs. When the model says "70% chance of failure," it should actually fail ~70% of the time.
Test data is never seen during training within each fold. Features are computed before outcomes are known. No future information leaks into predictions.
We used time-based validation to prove the model works on unseen data:
Have questions about the methodology, data, or accuracy?
Upload your CSV. Get AI predictions in 30 seconds. See pattern matches, similar companies, and hidden signals.
Required: Company name (that's it!)
Recommended: Sector + either (investment & value) OR (multiple)
Better predictions with: IRR, stage, revenue, employee count
We use smart defaults for missing data. More data = more accurate predictions, but even basic data gives useful signals.
Our proprietary feature extraction analyzes patterns that historically correlated with outcomes. Request detailed methodology →
Feature importance learned from training data (click for details)
These predictions are informational signals based on pattern matching—not financial advice. Always combine with your own due diligence and judgment.
A 30-minute deep dive for VCs seriously considering the Founding Partner opportunity.
95.2% accuracy, 3,027 snapshots, the offer summary.
4-model ensemble, 40+ features, validation methodology.
Honest answers to skeptical VC questions.
Landscape analysis and our defensible moat.
3-year roadmap from risk tool to intelligence platform.
Equity structure, expectations, and terms.
Concrete benefits for founding partners.
How anonymized data sharing works.
Real-time portfolio analysis walkthrough.
Login to access the presentation materials.
Invalid access code
Don't have a meeting scheduled?
Request a presentation →Everything you need to know
Regular users: Your CSV is processed in memory only. Never stored. Not used to train our model.
Founding partners: Data is encrypted, can be anonymized, and you can request deletion anytime (keeping your equity).
Privacy: Your data would go to their servers. Control: They can change terms. Specialization: GPT is general-purpose; our model is trained specifically on VC outcomes.
Encrypted at rest (AES-256), encrypted in transit (TLS 1.3), isolated per partner, no human access to raw data. You can request compliance audits and full deletion anytime.
It's real, and here's why:
We use time-based validation: train on past snapshots, test on future ones. This mimics real use (you have past data, want to predict future). The model sees snapshots from 2015-2020, predicts 2021-2023.
The ±4.2% variance shows consistency: we ran 7 different train/test splits, all between 86.5%-98.9% accuracy.
Why it works: Patterns that precede failure (declining margins + specific burn trajectory + growth deceleration) are remarkably consistent across companies and sectors.
107 companies × 28 snapshots each = 3,027 data points.
Quality over quantity. Every company has verified outcomes from SEC filings. Every snapshot is a quarterly financial report with ~40 features.
Compare to: most VC firms have seen ~50-200 companies total. We have 15x more pattern data than any individual investor could accumulate.
It doesn't "know" — it recognizes patterns. We trained it on 107 companies with 3,027 quarterly snapshots AND their outcomes. It learned patterns like "companies with declining margins + accelerating burn + flat growth failed 89% of the time." It compares your portfolio to these learned patterns.
We have 3,027 data points vs your ~50-200. We're data-driven vs bias-prone. We analyze in 30 seconds vs hours. BUT you have context we don't. Best approach: Use us to flag concerns, then apply your judgment.
You receive: Equity stake, board seat, lifetime access, roadmap input, model trained on YOUR patterns.
What I need: Portfolio data (can be anonymized), validation, network distribution, financial support (structure flexible).
Required: Sector, stage, investment amount.
Helpful: Revenue, employees, burn rate, valuation.
Outcomes: What happened (active/exited/failed).
You can anonymize company names. We only need metrics.
Public data is biased toward companies that got press. Partner data gives us real financials, quiet failures, and accurate outcomes. Your data makes the model better → better model protects YOUR portfolio.
Step 1: You export portfolio data (anonymized if you prefer). We need: sector, metrics (margins, growth, burn), outcome.
Step 2: We add your data to our training set. Run the same pipeline. Takes 5-10 minutes.
Step 3: New model is validated, deployed. Typically accuracy improves 2-5%.
Step 4: Everyone using the platform benefits. Your data contribution → your equity value increases.
Technical: Same Python pipeline (scikit-learn + XGBoost). No manual work. Automated validation ensures no accuracy regression.
The code? Yes. The model architecture is standard ML.
The data? No. Our moat is the network effect:
Bloomberg/PitchBook could theoretically build this, but they'd need to convince VCs to share private data. We're offering equity for data—they won't.
Yes. Request deletion anytime — we remove everything. You keep your equity. Product access continues. We retrain without your data.
We provide signals, not guarantees. Our predictions are one input for your decisions—not financial advice. 95.2% accuracy means ~4.8% are wrong. Always combine with your own due diligence, board conversations, and judgment. We're transparent about limitations so you can use this tool appropriately.
Still have questions?
amaan@portfolioray.com