Loading...
Loading...
Research
Honest research on what works, what doesn't, and what we learned building autonomous trading systems.
185 signals, 20.7% win rate, -389% cumulative P&L. Our AI trading bot was a disaster — until we discovered a single bug that could flip its accuracy from 21% to 79%.
When BTC consolidates for 10–24 hours, high-beta altcoins quietly misprice relative to their hedges. Our C++ scanner finds them. When BTC breaks, they move 2x as fast — generating 40–60% annual returns.
Most trading bots are glorified if-then-else machines. We built a system where 9 named AI agents reason, consult each other, and make collective decisions — with a human kill switch on Telegram.
We built a 75-feature ML pipeline for crypto futures prediction. After walk-forward validation, feature selection by IC, and honest analysis of what survived — here is what actually predicts 12-hour forward returns.
Our Silver ATR Breakout strategy showed a Sharpe ratio of 5-7 in backtests. After fixing one line of code — a missing .shift(1) — it dropped to -0.99. Here's our checklist so this never happens to you.
We maintain a library of 80 trading theses. Each one is evaluated by a structured pipeline: thesis → hypothesis → quick backtest → parameter search → walk-forward → paper trading → deploy. Most die at step 2. Here is the full process.
We analyzed 7,290 trading days and found that on 88% of them, price dips at least 0.5% from the previous close. A simple limit order ladder captures +109.9 bps of execution improvement.
We built an AI agent that generates comprehensive fundamental analysis reports on equities. After running it on AAPL and NVDA, here is what it gets right, where it struggles, and how it compares to human sell-side research.
After 26 consecutive SHORT trades hit stop-loss during a market bounce, we built a reversal detection system. Individual signals were weak (60-67%), but combinations reached 70-93% accuracy.
Thesis: assets in the bottom 10% by 7-day return tend to revert upward. We ran 40 backtests across 4 market regimes using 10 parameter sets. Here is the honest assessment of what we found.
We ran 7,560 metric judgments across 8 models, 15 prompts, and 3 judges. Claude Opus scored 90.5% clarity by one judge and 45.7% by another — on the exact same responses.
We ran 200 Monte Carlo trials across 94 symbols for 7 strategies. One strategy at Sharpe 1.0 was pure luck (99th percentile). Another at Sharpe 1.07 was genuinely robust (70th percentile).
Our ML crash detector uses 43 features and a Random Forest model. The top predictor? SOLUSDT 10-minute volatility. SOL moves first, the rest of crypto follows 10-15 minutes later.
Our mean reversion strategy on 15-minute bars showed Sharpe 4.52 and +32% returns in the 2022 bear market. Our own guidelines say Sharpe above 3 is suspicious. Here's our honest assessment.
50+ data sources, 500+ crypto symbols, 5,000+ equities, 6 countries. We built QData to unify it all into a single API — with Parquet storage, survivorship bias prevention, and automated CAPTCHA solving.
