Guide & Glossary

What the numbers on this dashboard actually mean. Project K.E.M. runs a two-arm system: one arm asks “is something unusual happening?” and the other asks “which way is price likely to move?” These definitions explain each output in plain language.

Every concept below is a research measurement, not a recommendation. See the Technology page for the architecture behind them.

Model Prediction & Confidence Novelty Score Concordance Timescales

Arm 2 · Probabilistic Forecasting

Model Prediction & Confidence

Arm 2 is a supervised transformer trained on historical outcomes. For each stock and each timescale it outputs a probability that price will move up versus down over that horizon — for example, “0.68 probability of a positive move over the next 5 trading days.”

Confidence is simply how far that probability sits from a coin-flip (0.5). A prediction near 0.5 is an admission of uncertainty; one near 0.9 or 0.1 is a strong directional call. Confidence measures the model's conviction, not its correctness — a confident model can still be wrong, which is exactly why the Performance page tracks realized outcomes.

Arm 1 · Novelty Detection

Novelty Score

Every day, all of a stock's data — price/technical, fundamentals, news text, and its position in the sector graph — is fused into a single high-dimensional embedding. The novelty score measures how far today's embedding sits from that stock's own historical baseline of embeddings.

Concretely it is computed from distance-based detectors — a Mahalanobis distance from the historical distribution and an Isolation Forest anomaly score. A high novelty score means the stock looks meaningfully different from its recent self: a regime the model has not seen much of before. Novelty flags that something changed; it does not say whether the change is bullish or bearish — that is Arm 2's job.

Arm 1 × Arm 2

Concordance

Concordance evaluates agreement between the two arms: does the unsupervised novelty trigger line up with a confident supervised directional call?

The interesting cases are when they agree — an unusual embedding (high novelty) and a confident directional probability point the same way. Disagreement is informative too: novelty with a near-0.5 probability means “something changed, but the model can't tell you which way,” which is a signal to be cautious rather than confident.

Timescales

Arm 2 forecasts three horizons independently. A model can be informative on one and uninformative on another, so each is scored on its own.

next_day

1 trading day

The most reactive horizon — sensitive to short-lived news and momentum, and the noisiest to evaluate.

short_term

5 trading days

Roughly one trading week. Smooths out single-day noise while still capturing near-term moves.

long_term

20 trading days

Roughly one trading month. Reflects slower, structural shifts in a stock’s representation.

“Trading day” means a day the market is open, so weekends and holidays are skipped. Twenty trading days is about one calendar month.

Want the architecture behind these concepts?

See how the two-arm pipeline works

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.