Project K.E.M.

How It Works

Project K.E.M. fuses four data modalities into one stock embedding, then runs a two-arm system: an unsupervised novelty detector and a supervised probabilistic forecaster.

From Data to Forecasts

Three stages

Ingest
1
For each stock we collect four data modalities: market/technical data, financial statements, news text, and sector-graph connections.
Fuse & integrate
2
Modality-specific encoders turn each source into a vector; a fusion layer combines them and a temporal integrator gives the stock a single, time-aware embedding.
Two arms + concordance
3
Arm 1 scores novelty against the stock’s baseline; Arm 2 forecasts directional probability. A concordance engine reports where they agree and publishes the result.

Pipeline

The two-arm architecture

Technical EncoderMarket DataFinBERT Text EncoderNews ArticlesFundamental EncoderFinancial StatementsGraph EncoderSector ConnectionsFusionLayerTemporalIntegratorArm 1 · NoveltyMahalanobis / Isolation ForestArm 2 · ForecastingSupervised TransformerConcordance EngineSignals & forecasts

Four modality encoders → fusion layer → temporal integrator → Arm 1 (novelty) and Arm 2 (forecasting) → concordance engine. See the Guide for what each output means.

Encoding each modality

Each data source is handled by a specialized encoder. Technical data flows through a time-series encoder over a canonical, fixed set of price and indicator features. Financial statements pass through a fundamental encoder; news text is embedded with a FinBERT-based text encoder; and sector relationships are encoded by a graph convolutional network. Every encoder emits a vector in a shared representation space so the modalities can be compared and combined.

Fusion & temporal integration

A fusion layer merges the per-modality vectors into one embedding, and a temporal integrator conditions it on recent history so the representation is time-aware rather than a single-day snapshot. The result is one embedding per stock per day — the common input both arms consume. Consistency here matters: every producer (bulk import, the daily pipeline, and model training) builds the stack the same way so embeddings from different processes live in the same vector space.

A note on proprietary bounds

We describe the mechanics openly for research transparency, but specific network dimensions, layer counts, and training hyperparameters are kept high-level. What matters for evaluating the dashboard is not the exact architecture — it is whether the published, forward-tracked predictions hold up against realized outcomes on the Performance page.

Explore the research

Read the concept definitions, browse the embedding space, or check realized performance.

Guide & GlossaryEmbedding SpacePerformance

IMPORTANT DISCLAIMER:

The information provided by Project K.E.M. is for informational purposes only and does not constitute financial advice, investment recommendations, or an offer to buy or sell securities. The signals generated are the result of a mathematical model and should not be the sole basis for any investment decision.

Past performance is not indicative of future results. All investments involve risk and the potential for loss. Users should conduct their own research and consult with qualified financial professionals before making investment decisions.

Project K.E.M. is not a registered investment advisor and does not provide personalized financial advice.

Academic research platform — not investment advice.