Overview
The AI Power Thesis
AI data centers require massive amounts of 24/7 carbon-free baseload power — a demand profile that intermittent renewables cannot reliably serve. Two companies sit at the intersection of this structural shift: Bloom Energy (BE), a commercial fuel cell provider with proven deployments, and Oklo Inc. (OKLO), a pre-revenue nuclear SMR developer with optionality on the long-duration clean power thesis.
This case study walks through VinePeak's full quantitative research pipeline applied to both stocks — from raw market data and synthetic news generation to embedding geometry, regime detection, and probabilistic 3-year forecasts.
Methodology
Research Pipeline
Six stages transform raw market data and information signals into probabilistic investment intelligence.
Stage 01
Market Data
5-year weekly OHLCV for target + 6 sector peers via yfinance. 11 market features: returns, momentum, volatility, relative strength, drawdown.
Stage 02
Synthetic News (LHS)
Regime-conditioned headlines generated via Latin Hypercube Sampling — ensures diverse sentiment coverage without over-representing any region of the signal space.
Stage 03
Text + Market Embeddings
Headlines → all-MiniLM-L6-v2 (384-dim, L2-normalised). Market features → PCA(10) + Johnson-Lindenstrauss projection → unit sphere.
Stage 04
Concentration Analysis
Lévy's Lemma bounds anomaly threshold at ε = 0.139 in 384-dim space. Rolling 26-week deviation scores flag genuine regime shifts vs. noise.
Stage 05
HMM Regime Detection
Gaussian HMM (4 states) on PCA-reduced embedding trajectory. Viterbi decoding labels regimes by return × volatility profile.
Stage 06
Monte Carlo Forecast
5,000-path regime-switching GBM with Bayesian-shrunk (James-Stein) parameters. Markov regime transitions from fitted HMM.
Mathematical Foundation
Concentration of Measure
In high-dimensional space, random noise self-cancels by the geometry of the sphere. Lévy's Lemma provides a rigorous bound: any 1-Lipschitz function on the unit hypersphere Sd-1 concentrates exponentially tightly around its median.
P( |f(X) − Median(f)| > ε ) ≤ 2 · exp( −(d−1) · ε² / 2 )
At d = 384:
ε = 0.125 → P < 0.02% // 90% anomaly threshold
ε = 0.139 → P < 0.003% // 95% threshold (used in this study)
ε = 0.166 → P < 0.00003% // 99% threshold
Any weekly embedding deviation beyond ε = 0.139 is not noise — it reflects genuine structural change in the information landscape. These anomaly weeks become the primary signal for regime transition detection.
384
Ambient embedding dimension
Both text and market embeddings projected onto S383
0.139
Lévy ε threshold (95%)
Deviations beyond this are statistically anomalous
~4.2
Effective intrinsic dimension
Low-dim signal structure from 384 ambient dims (both stocks)
60/40
Fusion weights
Text embeddings (60%) + market embeddings (40%) on unit sphere
~10%
Anomaly rate (both stocks)
Weeks with genuine structural signal vs. concentrated noise
5,000
Monte Carlo paths
Convergence by concentration of measure over paths
Market Data
Stock Profiles
Key market statistics and model outputs for each stock as of May 2026.
5-Yr Ann. Return
+50.5%
5-Yr Ann. Volatility
82.1%
Historical Sharpe
0.55
Current Regime
Funding_Winter
Walk-Forward Signal
BEARISH
Signal Accuracy
81.2%
Intrinsic Dim (emb.)
4.5
Anomaly Rate
10.0%
2-Yr Ann. Return
+100.7%
2-Yr Ann. Volatility
117.7%
Historical Sharpe
0.81
Current Regime
Regulatory_Setback
Walk-Forward Signal
N/A (short history)
Signal Accuracy
—
Intrinsic Dim (emb.)
3.8
Anomaly Rate
10.9%
HMM Output
Regime Analysis
Gaussian HMM fitted on the PCA-reduced fused embedding trajectory. Regime parameters calibrated with Bayesian shrinkage (James-Stein style) to prevent overfitting to short in-sample periods.
Bloom Energy — Detected Regimes
| Regime |
Weeks Detected |
Ann. Return (shrunk) |
Ann. Vol (shrunk) |
Signal |
| AI_DataCenter_Demand |
62 |
+80.0% |
50.1% |
Bullish |
| Normal_Growth |
10 |
−50.0% |
47.7% |
Neutral |
| Funding_Winter ← current |
90 |
+80.0% |
80.0% |
Mixed |
| Macro_RiskOff |
73 |
−50.0% |
64.1% |
Bearish |
Oklo — Detected Regimes
| Regime |
Weeks Detected |
Ann. Return (shrunk) |
Ann. Vol (shrunk) |
Signal |
| Nuclear_Renaissance |
26 |
+80.0% |
80.0% |
Bullish |
| Normal_Development |
5 |
−50.0% |
48.4% |
Neutral |
| Regulatory_Setback ← current |
30 |
+41.4% |
80.0% |
Caution |
| Capital_Crunch |
20 |
−50.0% |
80.0% |
Bearish |
Probabilistic Forecasts
Monte Carlo Scenario Analysis
5,000-path regime-switching simulation. Each path samples weekly returns from the current regime's calibrated GBM, with Markov transitions governed by the HMM transition matrix. Bar widths scaled relative to the Boom scenario.
Investment Advice
Recommendations
Position sizing follows Kelly half-fraction principles, calibrated to walk-forward signal accuracy and current regime transition probabilities.
Moderate Buy
Bloom Energy (BE)
Clean Energy / Fuel Cell · Current: Funding_Winter regime
- Core thesis: AI data center 24/7 baseload power demand is a durable, multi-year structural tailwind for commercial fuel cells
- Entry: accumulate on weakness toward P25 scenario (~$213); current regime flags short-term caution
- Walk-forward signal model achieves 81.2% directional accuracy — high confidence in signal quality
- Risk: hard stop below P5 (~$119); reduce on confirmed regime shift to Macro_RiskOff via embedding signal
- Kelly half-fraction sizing given vol of 82% and current bearish near-term signal
Suggested portfolio weight
3–5%
Speculative Buy
Oklo (OKLO)
Nuclear SMR · Current: Regulatory_Setback regime
- Core thesis: nuclear is the only scalable 24/7 zero-carbon source; Oklo is the pure-play pre-revenue expression of the SMR megatrend
- 3-year base case of +220% reflects genuine optionality, but binary regulatory path warrants small sizing
- Intrinsic embedding dim of 3.8 (vs BE's 4.5) signals thinner information coverage and higher idiosyncratic risk
- Structure: consider defined-risk options rather than equity, given extreme 3-year outcome kurtosis (skew +8.25)
- Monitor NRC licensing milestones — an adverse Information Request triggers embedding shift to Capital_Crunch → reduce
Suggested portfolio weight
1–2%
Portfolio Pairing
BE + OKLO: Complementary AI Power Exposure
BE provides near-term commercial revenue and a cleaner risk/reward profile. OKLO provides high-convexity optionality on the long-duration nuclear bet. The two are partially correlated through the macro energy narrative but diverge on regulatory and technology risk — making them additive within a concentrated clean-energy sleeve.
:
BE — commercial revenue, lower binary risk
OKLO — optionality
Combined clean energy sleeve: 4–6% of total portfolio, sized at Kelly half-fraction given ~66% average directional accuracy across both signal models. Rebalance quarterly or when the embedding regime signal changes.
What This Demonstrates
The VinePeak Method in Practice
This case study illustrates how each layer of the pipeline adds information that traditional analysis cannot access:
Geometry over statistics
Embedding deviation scores identify regime transitions before price fully reflects them — using the geometry of S383, not just return z-scores.
Honest uncertainty
Monte Carlo distributions — not point estimates. Bayesian shrinkage prevents overfit to short-sample regimes. The full probability fan is the output.
No lookahead bias
The walk-forward Ridge signal model trains only on past data at each step. 81.2% directional accuracy for BE is a genuine out-of-sample number.
Kelly-rational sizing
Position sizes derived from signal accuracy and regime volatility via Kelly half-fraction — not arbitrary conviction weights or equal weighting.
Disclaimer: This report is generated using synthetic news data, quantitative models, and publicly available market data. It is produced for research and educational purposes only and does not constitute financial advice, a solicitation, or a recommendation to buy or sell any security. All Monte Carlo forecasts are probabilistic distributions, not guaranteed outcomes. Past model performance and historical returns do not guarantee future results. VinePeak makes no representation as to the accuracy or completeness of the information herein.