How our engine generates signals

8-feature logistic model with stochastic gradient descent (SGD) and L2 regularization (ridge penalty). Numerically stable sigmoid with branch x≥0 vs x<0 to avoid overflow. Prior initialized from existing confidence formula coefficients, then updated with each labeled outcome. Proxy outcomes (0.3 weight) and real outcomes (1.0 weight) trained jointly.

The model learns continuously from tracked outcomes, adjusting its weights as new data arrives.

Custom 4×4 Gauss-Jordan matrix inversion for Hessian. Ridge regularization λ=0.5 to prevent coefficient explosion on sparse conflict data. Backtracking line search for step size. Converged in 7 iterations on 2,134 country-year observations with 4.55% base rate.

The same class of optimization used in epidemiological models and political science research, adapted for our conflict-onset predictions.

Train on 2010–2018 (1,746 observations, 3.32% base rate). Holdout 2019–2022 (776 observations, 6.19% base rate — period shift due to COVID + Myanmar + Sudan). Refit Brier = 0.0587, baseline Brier = 0.0591 (ratio 0.9929). Confirmed that v0 baseline transfers robustly despite base rate doubling.

Walk-forward validation prevents overfitting by testing on data the model has never seen during training.

Continuous monitoring of model Brier scores for distributional shift. Parameters: δ=0.005, λ_watch=0.05, λ_alert=0.10, λ_critical=0.20, minObs=20. Drift score = cumulative sum minus running minimum. Three severity levels trigger different responses.

Originally an industrial process control method, applied here to detect when market regime changes invalidate the model's assumptions.

Beta(α,β) posterior per (method, source, analyst) × event class. Update: Brier loss → α += (1-brier), β += brier, with exponential decay γ = 2^(-Δt/halfLife), halfLife=30 days. Naturally weights recent performance over stale predictions.

A probability distribution over reliability that accounts for recency, sample size, and domain — updated after every resolved prediction.

30-day rolling Pearson correlation between BTC daily returns and DXY daily returns. Confirmed inverse correlation used to adjust the Setup Score when DXY moves >0.5% in a session. Per-coin beta computed as Cov(coin,BTC)/Var(BTC) over rolling window.

Measures the live statistical relationship between crypto and the dollar, and adjusts the Setup Score accordingly.

New parameter sets run in shadow mode alongside the active model. Promotion criteria: shadow Brier < active Brier × 0.95 on ≥30 common resolved predictions. Demotion if shadow Brier > active × 1.10. Full audit trail with ParameterSet versioning, timestamps, and metrics snapshots.

A/B testing for prediction models — ensures new parameter sets prove themselves on live data before replacing the active version.

648 directed border edges between 250 countries (from REST Countries API). Bad neighborhood = 1 if ≥4 neighbors have UCDP fatalities ≥25 in adjacent years, or fallback WGI Political Stability ≤ -1.0. Batch pre-computation for O(1) lookup during prediction.

Graph-based spatial features drawn from conflict prediction research (ViEWS, PITF), implemented for country-level risk assessment.

f* = (p × b - q) / b where p = empirical win rate, b = avg_win / avg_loss, q = 1-p. Quarter Kelly (0.25×f*) applied for variance reduction. Setup-Score-scaled: maps the score [35,95] to multiplier [0.7, 1.3]. Risk level discount: MEDIUM ×0.75, HIGH ×0.5. Market cap safety caps: small-cap max 1x, mid-cap max 2x, large-cap max 5x.

Mathematically optimal sizing for long-term capital growth, with quarter-Kelly dampening to reduce variance.

Before any engine change ships, the full decision pipeline is replayed against a frozen library of captured real-market cases (inputs pinned, outputs byte-compared). Any unintended change in emitted signals, Setup Score, entries, TP/SL geometry or rejection reasons fails the gate. New behavior is only allowed behind explicit, fingerprinted configuration flags. An adaptive drawdown circuit breaker also exists in the codebase but is currently disabled by operator decision — every qualifying setup is published unfiltered and tracked publicly.

The same regression discipline used in flight software: no engine change reaches production unless it reproduces the frozen baseline exactly, or declares its behavioral change explicitly.

Hourly OI history from Binance futures (8 data points). Cascade detected when OI drops >5% in 4 hours. Direction classified by concurrent price move: OI down + price down = long liquidation (contrarian LONG signal); OI down + price up = short squeeze. Adjustment ±5 to derivatives score.

Identifies moments when leveraged positions are being forcefully closed, which often marks short-term exhaustion points.

Pool Adjacent Violators builds a monotone map from score to historical outcome frequency, refreshed from closed signals. We tested whether that map PREDICTS out-of-sample, and it does not: on purged, day-clustered holdout it scored WORSE than simply predicting the base rate (ΔBrier vs a constant −0.0130, 95% CI [−0.0208, −0.0052]). An independent affine re-test agreed (CI [−0.0235, +0.0020] — indistinguishable from a constant). The reason is that the score-to-outcome relationship is not stable across days: the same 55–65 band delivered 3% on one day and 66% on another. So the map is kept as a DIAGNOSTIC of calibration drift and is deliberately NOT used to convert the Setup Score into a win-rate.

We checked whether "score 70" reliably means a fixed win-rate. It does not — it moves with the market from day to day. So we do not publish one. The Setup Score ranks setups; the track record reports what actually happened.