Research

Ideas BlackNull is studying. Everything on this page is research, not a live product: no performance figures are shown because none has been verified yet.

Joint-embedding models for portfolio optimisation and asset discovery

A Joint-Embedding Predictive Architecture learns by predicting the representation of the missing part of an input from the visible part, instead of predicting raw values. For markets this matters: prices are noisy, so a model that forecasts every tick wastes capacity on noise, while one that predicts in a latent space keeps the stable structure (regime, risk factors, cross-asset links) and drops the rest.

How BlackNull would use it

1. Context encoder. Takes a window of history for many assets at once: prices, volumes, funding and borrow rates, filings, on-chain flows. It outputs one embedding per asset and one for the whole market.
2. Predictor. From the embeddings of the past window, predicts the embedding of a future window. The target encoder is a slow-moving copy of the context encoder (no pixel- or price-level reconstruction). The loss is the distance between predicted and actual embeddings.
3. Discovery. Assets whose embeddings move toward the region where good risk-adjusted outcomes lived in the past are flagged as candidates. Candidates are then checked by BlackNull's Risk Analysis before anyone sees them.
4. Reinforcement learning on top. A policy reads the frozen embeddings, not raw prices, and chooses portfolio weights. Its reward is risk-adjusted: return minus a penalty for drawdown, turnover, fees and concentration. Training in latent space is cheaper and less prone to over-fitting single price paths.
5. Hard limits outside the model. Position caps, minimum liquidity and the risk-score bar are rules in code that the policy cannot override.

What would have to be true before it goes live

• Walk-forward backtests on data the model never saw, with fees and slippage included, published with the code.
• A paper-trading period with real quotes and no money.
• Results compared with simple baselines (equal weight, buy and hold). If it does not beat them after costs, it is not offered.

Status: design only. No model has been trained and no result exists.

Pools from 13D filings

A Schedule 13D is filed with the US SEC when an investor takes more than 5% of a listed company and may seek to influence it. It shows what a large, informed holder chose to concentrate on, within days of the purchase. (A 13F, by contrast, lists a fund's holdings once a quarter, up to 45 days late.)

The plan

1. Read new 13D filings from the SEC's public filing database, public data.
2. Group them into pools (for example by sector or by the filing investor's track record), each with a stated rule for weights.
3. Hold the stocks as tokenised shares through a regulated provider (Tokenized treasuries Global Markets), inside a Multi Fund pool next to the existing ones.

Open points

• Who holds the assets: a pool needs one controlled address, kept by the platform and not by a person, ideally a multi-signature wallet. This is being decided.
• Eligibility: tokenised US shares are usually not offered to US persons and need identity checks in many places. The pool must follow the provider's terms.
• Delay: a 13D appears after the buy, so a pool copies late. Whether that still adds value has to be tested like any other strategy.

Status: not live. No pool is open and no deposit is accepted.