How do our Models Work?
Afternoon,
Every ticker that hits the dashboard has to survive a two-stage gauntlet: a raw-data ingestion layer (think 10-Ks, FDA dockets, satellite counts, patent filings) and a second layer of bespoke risk engines tuned to the lifecycle of the company. Below is the plumbing behind each sleeve, no marketing gloss, just the math we run before your capital goes to work.
Revenue Model – “Early-Cash-Flow Engine”
Universe: 2,800 U.S. listed names with ≥ $20 m TTM revenue and ≤ $5 b EV
Core signal stack
a. Revenue-Quality Vector (RQV) – 9-factor z-score
• Recurrence: % revenue that is contracted or subscription
• Cash conversion: ΔOFCF / ΔRevenue
• Unit-economics slope: (GM$ per customer) / (CAC)
• Revenue-to-R&D elasticity: how much incremental revenue is generated per $1 of R&D
We require z ≥ 1.5 on the composite; anything < 0 is auto-rejected.
b. Growth-Sustainability Filter – stochastic frontier regression
We estimate a “maximum efficient growth” frontier given balance-sheet capacity. Companies growing > 25 % above the frontier are down-weighted 30 %; those below the frontier get a 15 % boost (less burnout risk).
c. Downside-Capture Shield – spectral tail-risk model
A 3-factor GARCH-DCC copula simulates 10,000 21-day paths; we cut the equity if 95 % CVaR > –18 % in any rolling quarter.
Position sizing
Risk-budget = (RQV × 0.4) + (inverse CVaR × 0.4) + (liquidity score × 0.2)
Max single name 8 %, max sector 25 %. Rebalance: monthly, or intra-month if RQV drops below 1.0.
Live metrics (close 9/26)
6 holdings | +89.6 % unrealized | Sharpe 2.1 | Max DD –11.4 % | Beta 0.88
Pre-Revenue Model – “Catalyst-to-Cash-Flow Converter”
Universe: 650 pre-commercial equities (pre-IPO SPACs, biotech Phase 1–2, space, quantum, fusion)
Core signal stack
a. Path-to-Cash-Flow Tree – ensemble of 1,200 venture-funded comps
Random-Forest with 150 features: CTO pedigree, patent forward-cites, grant dollars, hiring velocity, addressable-market elasticity. Output is probability of reaching positive FCF inside 36 mo. We require p ≥ 0.35; median in universe is 0.18.
b. Technology Viability Index – graphed patent network
Page-rank algorithm on USPTO citation graph; we want eigenvector centrality ≥ 70th percentile within the sub-sector.
c. Financing-Runway Monte-Carlo
Burn, dilution, and milestone payments are simulated 5,000 paths. If Pr(cash-out) > 30 % inside 18 mo, position is capped at 2 %.
Position sizing
Kelly-fraction scaled by 0.25 (we never take full Kelly). Liquidity taper: if 20-day ADV < 75 k shares, weight is haircut 40 %.
Live metrics
3 holdings | +169.9 % unrealized | Sharpe 3.4 | Max DD –19.7 % | Beta 1.42
Dividend Model – “Coverage-over-Yield”
Universe: 1,100 dividend-paying U.S. equities, REITs, BDCs, MLPs
Core signal stack
a. Dividend-Coverage Probability – dynamic probit model
Explanatory vars: FCF payout, interest-coverage, EBIT margin volatility, tax-rate delta. We require p(cut) ≤ 10 % over next 12 mo.
b. Growth-at-Reasonable-Yield (GARY) – penalized regression
Target: 10-yr dividend CAGR. Penalty term for starting yield > 7 % (value trap filter). Only stocks with predicted CAGR ≥ inflation + 200 bp survive.
c. Rate-Shift Hedge overlay
Duration of dividend stream is computed via a 2-factor Vasicek model. If effective duration > 7, we pair the equity with a micro Treasury short position inside the sleeve (not the account) so net duration < 3.
Position sizing
Equal-weight at rebalance, but if trailing 12-mo volatility > 25 %, weight is reduced pro-rata to keep sleeve volatility ≤ 15 %.
Live metrics
5 holdings | +4.3 % unrealized | Yield on cost 4.8 % | Sharpe 1.0 | Max DD –6.9 % | Beta 0.65
Micro/Small-Cap Pharma Model – “Binary-Event Engine”
Universe: 420 North-American listed biotechs with market cap ≤ $2 b and at least one asset in Phase I–III
Core signal stack
a. Clinical-Probability of Success (CPoS) – Bayesian hierarchical model
Training set: 3,700 historic trials. Covariates: mechanism of action, biomarker status, trial design, FDA precedent, prior probability adjustment from peer-reviewed literature. Posterior CPoS is updated within 30 min of any public data release.
b. Regulatory-Insight Parser – fine-tuned BERT on 18 k FDA correspondence docs
Flags linguistic sentiment of FDA minutes, CRLs, SPA letters. A z-score < –1.5 triggers an immediate position review.
c. Market-Opportunity Elasticity – copula of TAM vs. peak-sales consensus
If modeled peak-sales < 25 % of TAM, we assume pricing pressure and haircut NPV 20 %.
Position sizing
Kelly/4 with a hard 50 % stop-loss from cost. Liquidity filter: 30-day median dollar-volume ≥ $1 m; else position is rejected.
Live metrics
8 holdings | +71.8 % unrealized | Hit rate on binary events 68 % | Sharpe 2.6 | Max DD –22.3 % | Beta 1.15
Cross-Model Risk Governance
• Weekly covariance overlay: we run a 250-day dynamic conditional correlation (DCC) model across all sleeves. If predicted 30-day portfolio VaR > 8 % at the 95 % level, we shrink the two highest-beta sleeves (usually Pre-Rev and Pharma) by 20 %.
• Single-factor crowding check: we compare our aggregate gamma (options-adjusted delta) to the Bloomberg U.S. Equity Crowding Index. If overlap > 30 %, we trim.
• Draw-down protocol: at –12 % peak-to-trough for any sleeve, we automatically move 30 % of that sleeve’s capital into 3-month T-Bills until the drawdown recovers to –6 %.
What You’re Missing
The numbers above are point-in-time. The models retrain every weekend after Friday’s close: new 10-Ks, trial uploads, patent grants, and options flow are ingested by Sunday 6 p.m. ET.
We’ll see you at the next rebalance.