About
About AlgoZilla
We build algorithmic trading systems that survive contact with reality. No marketing-optimized backtests. No cherry-picked results. Just rigorous, walk-forward validated signals.
Our Story
AlgoZilla started from a simple frustration: most trading signals are worthless. They look amazing in backtests but fail miserably in live markets. The reason? Overfitting, lookahead bias, and curve-fitting to historical data.
We set out to build something different. A system that uses the same rigorous walk-forward validation methodology employed by institutional quantitative funds — but accessible to individual traders.
After 78 iterations of development, extensive A/B testing, and thousands of hours of compute time, AlgoZilla delivers trading signals with genuine out-of-sample edge. Every metric we publish comes from walk-forward validated results — never in-sample backtests.
Our Philosophy
One Change at a Time
We never combine untested changes. Each improvement is validated individually before being added to the production pipeline. This scientific approach ensures we know exactly what works and why.
Investigate First, Conclude Later
When something goes wrong, we check the data before drawing conclusions. No assumptions. No shortcuts. Logs, metrics, and evidence drive every decision.
Transparency Over Marketing
We show all metrics — including the ones that don’t look great. Bull regime performance, drawdown periods, losing streaks. If you’re going to trust a signal, you deserve to know its weaknesses.
The Technology
Built on proven quantitative finance infrastructure, running on dedicated high-performance servers.
Ensemble Classifiers
Three independent machine learning models (LightGBM, XGBoost, GradientBoosting) must agree before a signal is generated. This consensus approach dramatically reduces false positives.
Regime Detection
Markets cycle through Bull, Sideways, Bear, and Bubble phases. Our EMA-based regime detection system automatically adjusts trading parameters — stop-losses, trailing multipliers, and entry thresholds — for each phase.
26,000+ Features
From Lorentzian distance classifiers to Ichimoku cloud analysis, from fair value gaps to harmonic patterns. Information Criterion filtering selects only features with genuine predictive power — typically ~110 out of 26,000.
Walk-Forward Testing
6-fold expanding window with embargo periods between train and test sets. Every result we publish is truly out-of-sample — the model never saw the test data during training.
Dedicated Infrastructure
Our signals run on dedicated bare-metal servers with 512GB+ combined RAM and 72+ CPU cores. No shared cloud instances. This ensures consistent signal generation with no delays.
Continuous Validation
Paper trading runs continuously to verify that walk-forward results hold in live market conditions. Monthly reports compare paper trade alpha against buy-and-hold benchmarks.