Omniscius — v6.2 (Pythia)
Three years of quantitative research. One audited code path: every live signal is produced by the exact same logic stress-tested in backtest — proven bit-equivalent. Independently validated, leak-audited, reproducible. The v6 release introduces principle-driven feature engineering: statistical evidence decides which signals matter, not arbitrary thresholds.
Pythia v6.2 is the latest iteration of the Omniscius machine learning engine that powers every AlgoZilla trading signal. It represents three years of quantitative research, backtesting, and iterative refinement — from early prototypes with basic indicator crossovers to a sophisticated ensemble classifier that lets statistical evidence decide which signals matter at each market moment.
The Greek name Pythia refers to the Oracle of Delphi — the trusted source consulted before every major decision. That captures the v6 philosophy: trust the data. Instead of arbitrary correlation cutoffs and feature caps, v6.2 lets per-coin permutation tests decide which features survive into the live classifier. Significant features pass through. Pure noise gets filtered out. The model never sees the difference.
What sets the engine apart is its methodology. Every signal is generated through walk-forward validation: train on the past, test on unseen future data — never use the same data twice. The v6 release builds on three prior generations of research:
- ✓ Itzamná (v4.0/v4.0.1) confirmed the statistical edge across 46 coins via independent null-tests, then unified backtest + live into one bit-equivalent code path.
- ✓ Katharsios (v4.1) restored the calibrated hold-zone in sideways markets.
- ✓ Theros (v4.2) liberated the cooldown for selective re-entry.
- ✓ Kairos (v5.0) extended that liberation to all five regimes and more than doubled the feature library.
- ✓ Pythia (v6.0 → v6.2) introduces data-driven feature selection across 7,700+ candidate features, principle-driven filtering, and signal-aware stop logic.
The model is retrained per coin on a regular cadence. Market dynamics change — new correlations emerge, old ones decay. Continuous retraining ensures the engine adapts to current conditions rather than relying on stale patterns.
We believe in continuous improvement. Every iteration is rigorously tested against the previous version before deployment. We never ship changes that don’t demonstrably improve walk-forward performance.
How It Works
7,700+ Candidate Features
Every coin is analyzed across momentum, trend, volume, volatility, sentiment, microstructure, candle patterns, macro crosses, and proprietary regime indicators. The v6 library expanded dramatically with cross-family orthogonal interactions, monthly-anchored sentiment, synthetic regime detectors, and macro-lag features — all evidence-tested per coin before reaching the live classifier.
Explore indicators →Data-Driven Feature Selection
Instead of arbitrary thresholds (the legacy “drop features below IC = 0.04” approach), v6.2 lets per-coin permutation tests decide which features survive. Features that show statistically significant predictive power in at least one coin / horizon pair pass through. Features that look like noise across all coins get filtered out automatically. The model receives an evidence-curated input, not a hand-tuned one.
Pulse Regime Detection
Every coin gets its own Pulse classification per hour: INERT (capitulation), CHARGING (accumulation), AWAKENING (regime-pivot), RAMPAGE (bull run), or ATOMIC (distribution peak). Entry thresholds, exit sensitivity, and feature relevance all adapt automatically per regime. Live Pulse →
The Feature Library
A deeper look at what Pythia sees on every coin, every hour.
Momentum & Trend
RSI families (5 / 14 / 21 across 6 timeframes), MACD histograms, ROC ladders, Bollinger position, Hilbert phase & dominant cycle, ADX-strength. Every base indicator emits both level and delta variants — the engine sees not just where the market is, but how fast it got there.
Volume & Microstructure
VWAP-deviation, on-balance volume, accumulation-distribution, micro-extreme magnitude, post-shock decay, candle anatomy (body / wick ratios), 90+ candle-pattern signals from the TA-Lib library — including the proprietary Cherenkov suite (DeMark squared with radioactive half-life weighting).
Macro & Sentiment
Fear & Greed Index (multi-timescale), DXY / gold / silver / brent / VIX / TNX cross-products, BTC dominance, total market cap, on-chain narratives, monthly-anchored sentiment averages. v6 introduces R21 macro-lag features (with full forward-leak audit) and a synthetic Fear & Greed computed directly from price action when external data lags.
Cross-Family Interactions (NEW)
The R5h pack: 37 orthogonalized cross-family interaction features mixing macro × coin × volatility × liquidity × cycle. Plus R5b: 42 Fear & Greed × Pulse cross-products. These capture the patterns no single indicator can — e.g. how on-chain breadth interacts with VIX during BUBBLE regimes.
Pulse-Native Features
Per-coin Pulse classification drives a dedicated feature sub-pack: persistence indicators (how long has this regime lasted), velocity (rate of regime change), acceleration (second-order regime dynamics), and synthFG — a price-action-derived Fear & Greed proxy that works on every coin individually, not just the BTC-driven market average.
Statistical Validation
Every feature is independently permutation-tested per coin and per prediction horizon (16 horizons from 1h to 672h). A feature must show statistically significant predictive power before it enters the live classifier. Pure noise gets blacklisted automatically. No magic numbers, no arbitrary IC thresholds — evidence-driven by design.
Walk-Forward Validation
Walk-forward validation is the gold standard in quantitative finance. The concept is simple: train on the past, test on the future. Never use the same data twice.
Pythia is validated across 3 Bitcoin halving cycles. Each cycle represents fundamentally different market conditions — from the 2017 ICO mania to the 2022 bear market to the 2024 ETF-driven rally and beyond. A model that performs well across all three isn’t lucky — it has found genuine statistical edges.
For v6 we kept the proven blended d0/d1 mcap-Sharpe KPI from Kairos — it matches real live-execution latency (~30–90 seconds) far more honestly than the legacy 4-bar metric that quietly penalized real alpha.
We publish backtest metrics transparently: Sharpe ratio, win rate, profit factor, maximum drawdown, and number of trades. These aren’t cherry-picked — they’re averages across all validation folds.
Continuous Improvement
Pythia v6.2 is the result of hundreds of experiments. From v0.1 (a simple RSI crossover) through the v4 lineage (Itzamná statistical edge → Katharsios hold-zone → Theros cooldown liberation), v5.0 Kairos (all-regime cooldown release + 506-feature library), to v6 Pythia — the principle-driven generation.
Our research process follows strict scientific methodology: one variable per test, independent validation sets, and statistical significance testing before any change is accepted.
Every improvement is documented and version-controlled. When we deploy a new version, we publish the changelog below so you know exactly what changed and why.
Current Performance
*Walk-forward backtest results (2017–2026). Past performance does not guarantee future returns.
Changelog
Every significant model update is documented here. Minor retrains are not listed.
v6.2 (Pythia) — Principle-Driven Selection
Data-driven feature selection replaces arbitrary IC and correlation cutoffs. The “let evidence decide” philosophy: per-coin permutation tests determine which features survive into the live classifier. The candidate pool expanded to 7,700+ features — macro × sentiment cross-products, monthly-anchored Fear & Greed indicators, synthetic regime detectors derived from price action, and 37 cross-family orthogonal interactions. Signal-aware stop logic respects the model’s conviction: when a strong signal is still present, hard stops do not trigger prematurely. Every code path leak-audited and bit-equivalent across backtest and live execution.
External data sources integrated in v6
Pythia ingests live data far beyond OHLCV. The v6 release expanded the external data stack significantly:
- ✓ Macro markets: DXY, gold, silver, brent crude, S&P 500, Nasdaq 100, VIX, 10-year Treasury (TNX) — hourly resolution with full forward-leak audit on every lag computation
- ✓ Sentiment: Fear & Greed Index (Alternative.me) at multi-timescale resolution (raw + 30d / 90d / 180d / monthly anchors)
- ✓ Crypto-market structure: BTC dominance, total market cap, alts vs BTC breadth — synthetic hourly series back to 2017 (back-computed from base supply + spot prices)
- ✓ Derivatives signals (Sprint 1.5 expansion): Coinalyze futures open interest + liquidations, Binance klines + funding rates — per-coin where available
- ✓ DeFi & on-chain: DefiLlama TVL aggregates, Hyperliquid order-book metrics, CoinPaprika market structure
- ✓ Synthetic Fear & Greed: a price-action-derived Fear & Greed proxy (vol 29.4% + momentum 29.4% + breadth 17.6% + dominance 11.8% + trends 11.8%) for coins where the external index lags or is missing
Feature engineering on top of external data
- ✓ Per-coin permutation tests replace arbitrary IC / correlation cutoffs
- ✓ 7,700+ candidate features (was 506 in Kairos)
- ✓ Cross-family orthogonal pack: 37 macro × coin × volatility interactions
- ✓ Signal-aware stop logic: hard stops respect strong-signal conviction
- ✓ Monthly-anchored Fear & Greed features for regime persistence
- ✓ R21 macro-lag features (DXY / gold / silver / brent / SPX / NDX / VIX) with verified +1 day temporal alignment
- ✓ All feature pipelines leak-audited (full forward-leak elimination)
v6.1 (Pythia) — Sprint 1.5 / 1.6 Feature Expansion
Major feature library expansion building on the v6.0 R2c foundation. The seven Sprint feature groups landed: R5b (42 Fear&Greed×Pulse cross-products), R5c (sentiment cross suite), R5d_v2 (monthly Fear&Greed anchors), R5f (Pulse EMA10 ladder), R5g (SynthFG promotion), R5h (cross-family orthogonal pack, 37 features) and R21 (macro-lag features). Legacy 661-feature blacklist removed in favour of evidence-driven selection.
v6.0 (Pythia) — R2c Foundation
Architectural reset and audit. Seven bulk-regex bugs fixed in the feature library, feature semantics corrected (ema-spread raw, micro-extreme magnitude cap), four EXO dead-code features brought back online. Foundation laid for the principle-driven feature selection that became v6.2.
v5.0 (Kairos) — Trust the Model, Everywhere
Cooldown set to zero across all five regimes. Three structural bugs in the BUBBLE scorer fixed. The feature library more than doubled to 506 features per coin per hour. Per-coin Pulse classification (INERT / CHARGING / AWAKENING / RAMPAGE / ATOMIC) replaces the legacy 4-regime system. Single-bar exit confirmation replaces the 2-bar lag.
v4.2 (Theros) — Cooldown Liberation
Replaced the legacy 12-bar cooldown lock with a 1-bar guard outside BULL. Combined with a tighter entry-price reference, validated on an 11-way apples-to-apples sweep: mcap-Sharpe +0.394, median +0.599, 12/12 coins materially better vs the Katharsios baseline.
v4.1 (Katharsios) — Hold-Zone Restored
Removed a legacy macro-override that was collapsing entry-and-exit thresholds onto a single point in sideways markets. The gridsearch-calibrated hold-zone is now respected in all regimes.
v4.0 + v4.0.1 (Itzamná) — Statistical Edge Confirmed + Unified Execution
Major architectural overhaul (v4.0): independent out-of-sample tests across 46 coins plus multiple null-tests (label shuffle, reverse-walk, OOS-replace) confirmed the statistical edge is real. Forward-looking-bias audit fixed 40+ subtle leaks. The v4.0.1 patch unified backtest and live trading into one shared decision module — every entry filter, every exit rule, and every threshold identical across both paths. Proven bit-equivalent: 0 trade-level mismatches over 5,318 trades, 99.984% per-bar agreement on a full BTC walk-forward.
- ✓ Statistical edge confirmed via independent null-tests on 46 coins
- ✓ 40+ forward-looking-bias leaks eliminated
- ✓ Single shared decision module: backtest = live (bit-equivalent)
v2.5 — Production Release
Full ensemble classifier with LightGBM. Walk-forward validation across 3 halving cycles. Per-coin retraining. Real-time Telegram delivery.
v1.8 — Regime-Aware Thresholds
Introduced adaptive entry and exit thresholds based on market regime.
v1.5 — Multi-Timeframe Expansion
Added multi-timeframe analysis (2H, 4H, 8H, 12H, 24H) for all base indicators.
v1.0 — First Public Release
Initial public release with BTC, ETH, and XRP coverage.
v0.x — Research & Development
Two years of research, prototyping, and backtesting.
