Signals in Thin Air

A kNN lens on earnings-driven turbulence

Thin liquidity around holidays, discrete information shocks from earnings, and clustered volatility create a recurring microstructure: returns around like states rhyme. We exploit this with a minimalist k-nearest neighbours (kNN) classifier that maps current market state vectors to a historical neighbourhood and infers a short-horizon directional bias. The model is intentionally simple to keep the inductive bias legible; complexity is reserved for the features, not the learner.

State vector

For each symbol i, we form a monthly-to-daily fused state at end of day t:

• Momentum & curvature: z-scores of 5/10/20-day log-returns; discrete trend sign; quadratic fit curvature over 10 days.
• Volatility regime: EWMA vol (10, 20), realized quarticity, and a GARCH(1,1)-implied next-day sigma.
• Event proximity: business-day distance to scheduled earnings; binary window flags [-3, +3].
• Microstructure: % gaps, overnight/regular-session split, volume percentile, order-imbalance proxy.
• Cross-sectional context: beta to Nasdaq-100, semis factor, and residual idiosyncratic return.

Distance, neighbourhood, label

We use Mahalanobis distance on a whitened feature space (robust covariance) with an event-weighted kernel that downranks non-earnings neighbours when today is an earnings window. The label is sign(rt+1) (next session close-to-close), and we report class probability via neighbour vote proportions with temperature scaling.

Cross-validation & leakage controls

• Purged, embargoed, expanding-window CV to avoid look-ahead around events.
• Time-kNN baseline vs. feature-kNN ablation to attribute lift correctly.
• Post-earnings drift filter: require positive drift in neighbours when t is in [0,+2] post ER.

Signal construction

We emit a positive-bias flag when Pr[rt+1>0] ≥ 0.55, neighbour count ≥ k=21, and entropy < threshold. Position sizing (if any) is outside scope; this note is informational only.

Screen — 10 Nasdaq names with elevated event volatility (earnings window this week; model bias ≥ 0.55)

"Positive" here denotes that the model’s neighbour majority implies a >55% probability of a next-session positive close; it is not a recommendation. The list is intentionally diversified across sectors to avoid spurious factor concentration.

Why earnings windows amplify kNN signal

Around discrete disclosures, conditional distributions of returns are multi-modal and regime-dependent. A non-parametric learner like kNN can adapt to these local structures without imposing a Gaussian prior. When features encode event proximity and volatility regime, the neighbourhood collapses to a small manifold of historically analogous states (e.g., banks in a rising-rate cut probability environment, semis during AI capex cycles), improving calibration.

Pseudocode

// X: feature matrix (T × d), y: sign of next-day returns, today: x_t
k = 21
N = neighbourhood(X, x_t, metric = mahalanobis, kernel = event_weighted)
prob_up = mean(y[N] == +1)
if (prob_up ≥ 0.55 and |N| ≥ k and entropy(y[N]) ≤ τ) emit("positive-bias")
else emit("no-edge")