I
Point-in-time, always
Each year's universe, features and labels use only what was knowable on that date. Index membership is rebuilt from historical snapshots, so names that were later acquired appear and drop out on schedule.
Quantitative research · US large-cap equities
Finding the principal direction in a noisy cross‑section.
EigenLane Quant Labs builds point-in-time machine-learning models for selecting US large-cap equities. Our first model, Meridian, is the research foundation for a proposed long-only strategy, EigenLane Fund I, which has not launched yet.
StageResearch & validation
FundNot launched
Live track recordNone
RegistrationNot a registered adviser
InboundOpen to introductions
01 · Approach
Research standards come first. Returns come after.
A back-test is only as good as the discipline behind it. These are the rules we hold every model to before it goes anywhere near capital.
II
No shared future
Every annual model is trained walk-forward, with a gap of roughly 390 days between its last training date and its entry date. A 12-month purge keeps forward-return labels from leaking into validation.
III
Like-for-like benchmarks
An equal-weight top-200 universe and a 12–1 momentum screen go through the same engine as the model: same snapshots, same entry dates, same costs. The model has to beat simple alternatives on equal terms.
IV
Weakest defensible statistic
We quote the conservative figure: medians over means, cohort-clustered errors over daily ones, and results with the best year removed. When we know about a defect, we disclose it before anyone has to find it.
02 · Meridian
One decision a year, made the same way every year.
Meridian is a gradient-boosted classifier that ranks the largest US equities once a year and holds an equal-weight basket for twelve months. The design is deliberately simple: anyone can audit it from a list of tickers and a date.
-
01
Universe
The 200 largest S&P 500 constituents by market capitalisation, as they stood on each 30 June. Eleven point-in-time snapshots, 2016–2026.
-
02
Features
Fundamental, valuation and technical characteristics, rank-normalised across stocks on each date and pruned by correlation clustering. Change terms are computed on fundamentals only, so price momentum can't re-enter through the back door.
-
03
Model
LightGBM, trained walk-forward with cross-validation and a multi-seed ensemble. Hyperparameter tuning is kept only if it beats the untuned baseline under the same scoring protocol.
-
04
Portfolio
30 names, equal-weighted, entered in late June and held for twelve months, with a 20 bps one-way transaction cost assumed. One rebalance a year.
200names in the point-in-time universe each year
1,791candidate features before pruning
11independently trained walk-forward models
~390dgap between last training date and entry
30names held, equal-weight
1/yrrebalance, June entry
What we share under NDA
Enough for qualified parties to check our claims independently using their own data:
- Universe definition and every point-in-time snapshot
- Entry and exit dates and the 30-name list for every historical cohort
- Entry price, exit price and return for every historical position
- Daily NAV, risk statistics and factor attribution
- Validation methodology and our full limitations register
What we keep proprietary
Everything needed to verify the claim is disclosed. Everything needed to rebuild it is not.
- The composite feature framework and its construction
- The candidate feature library and the pruned working set
- Hyperparameters, training windows, seeds and ensemble configuration
- Source code for the data and modelling pipeline
03 · EigenLane Fund I Proposed · not launched
A long-only, higher-beta US large-cap strategy.
The proposed design is below. Vehicle, terms and jurisdiction have not been set and will be decided with counsel. Any offering would be made only through definitive documents, and only to eligible investors.
- Mandate
- Long-only US large-cap equities
- Signal
- Meridian (point-in-time ML selection)
- Universe
- Top-200 S&P 500 constituents by market cap
- Construction
- ~30 names, equal-weight
- Rebalance
- Annual
- Target market beta
- ≈ 1.2
- Leverage / shorting
- None
- Expected eligibility
- Accredited investors only (TBD)
- Structure
- Private fund, to be determined with counsel
Who this would not suit
- Beta above one. The strategy is designed to move more than the market, both up and down.
- Deep drawdowns. In back-testing, the portfolio fell about 35% from peak to trough in early 2020. Expect drawdowns of that size as normal, not exceptional.
- Concentration. About 30 names, rebalanced once a year, so it can lag badly in narrow, mega-cap-led markets.
- Long horizon. Each decision plays out over twelve months. The strategy can't show anything meaningful over shorter windows.
- No live record. Everything to date is hypothetical and back-tested.
04 · Path to launch
Staged against research milestones, not a calendar.
Meridian v1 research
Point-in-time universe, walk-forward engine and like-for-like benchmarks. Ten annual cohorts back-tested.
Robustness & out-of-sample
Sweeps across entry months and seeds, beta control, and tracking the 2026 cohort as live out-of-sample evidence.
Structure & counsel
Choosing the vehicle, the regulatory path (for example, as an exempt reporting adviser) and service providers.
EigenLane Fund I
Launch only once the evidence and the structure both support it.
05 · Team
Engineering rigour meets quantitative research.
We're a small team, so key-person risk and operational depth are real limits. We'll discuss both openly in diligence.
Engineering leader and long-time equity investor, shaping EigenLane's strategy and product and building the systems that turn research into an institutional-grade process.
- Drives strategy, product, engineering and operations for EigenLane in the US, including the point-in-time data pipelines, the walk-forward back-test engine and the audit trail behind every result.
- Over a decade investing in public equities and more than five years in private equity.
- VP Engineering at a GPU cloud company (stim.io). Former founder (TaggedWeb / AITextBox). Early senior engineer at GoodRx. Ex-Amazon, where he built petabyte-scale stream-processing pipelines that integrated models into production data flows.
- Passed the Series 65. MS Computer Science, Oregon State University (Intel–OSU Graduate Fellow). B.Tech Electronics, IIT (BHU) Varanasi.
Derivatives and quant strategist who has taken a machine-learning stock-selection model from research all the way to an investable product.
- Leads model development for Meridian: feature research, walk-forward training and validation.
- Associate, Derivatives Strategy & Quant at Nomura (2022–2026): quantitative research and strategy used by internal traders and by external hedge fund and CTA clients.
- Previously built a machine-learning stock-selection model that was turned into an Actively Managed Certificate.
- In-house quant and cross-asset research for a $250m+ multi-asset fund at Zeus Knowledge Center; long/short equity research at GNY Capital.
- MS Finance, University of Maryland (Smith). B.Tech Electronics, IIT (BHU) Varanasi.
06 · Register interest
The right next step is a conversation, not an allocation.
We welcome introductions from allocators, family offices, quantitative desks and data buyers. If there's a fit, we'll share our research materials under NDA, including the full selection history, per-position returns and our limitations register.
Registering interest creates no obligation on either side, and doesn't make you a client or investor. We don't accept capital through this site.