Methodology

How the Bayesian fusion engine works

Methodology

1. Data Pipeline

• 10 cryptocurrency pairs on the MEXC exchange (BTC, ETH, SOL, BNB, XRP, DOGE, PEPE, WIF, SUI, AVAX)
• 5 timeframes: 1m, 5m, 1h, 4h, 1d
• OHLCV ingestion every 60 seconds
• PostgreSQL time-series storage with upsert deduplication

2. Strategy Layer

Eight independent strategies produce signals (BUY / SELL / NEUTRAL) with confidence scores in the range 0.0–1.0.

Technical strategies:

• RSI + MACD (momentum)
• Bollinger Bands (mean reversion)
• EMA Crossover (trend following)
• Volume Spike (volume confirmation)

Empirical strategies (academic evidence base):

• Intraday Time-Series Momentum (TSMOM)
• Post-Shock Mean Reversion
• Walk-Forward EMA with Volatility Targeting

Sentiment strategy:

• A composite score from the Fear & Greed Index, CryptoCompare news, crypto RSS feeds (11 sources), and macro feeds (FT, Fed, CNBC)

3. Market Regime Detection

Using EMA(200) + ADX(14) on daily candles:

• BULL: price > EMA200 and ADX ≥ 25
• BEAR: price < EMA200 and ADX ≥ 25
• SIDEWAYS: ADX < 25

The regime sets the Bayesian prior probability of an upward move, P(BUY):

• BULL → prior = 0.60
• BEAR → prior = 0.40
• SIDEWAYS → prior = 0.50

4. Bayesian Fusion Engine

The engine combines all eight strategy votes as a naive-Bayes product:

P(BUY | s1…s8) ∝ P(BUY) × ∏ P(si | BUY)^(2 × wi)

where wi is the reliability weight of strategy i, learned from historical performance. The exponent amplifies high-reliability voters and damps the rest.

Decision thresholds on the posterior confidence:

• confidence > 0.65 → EXECUTE
• confidence 0.35–0.65 → WATCH
• confidence < 0.35 → SKIP

5. Adaptive Daily Reweighting

Every night at 02:00 UTC:

1. Performance metrics are computed (24h and 7d win rate, Sharpe ratio).
2. Market regime is detected per symbol.
3. Regime multipliers are applied — momentum strategies are boosted in trending markets, mean-reversion strategies in sideways markets.
4. Claude AI (Anthropic claude-sonnet-4-6) analyses the performance data and recommends weight adjustments.
5. Final weights blend three sources: statistical (50%) + decay (30%) + AI (20%).
6. The weights are persisted and applied to the next day’s fusion.

6. Shadow Portfolio

A $1,000 USDT paper-trading portfolio tracks system performance:

• Position sizing: 1% of portfolio risk per trade
• Stop loss: 1.5× ATR below entry
• Take profit: 3× ATR above entry (2:1 reward/risk)
• Transaction costs modelled: MEXC 0% maker, 0.1% taker
• No real money is traded

7. Blog Generation

Daily Alpha Brief posts are generated automatically:

• Data is extracted from the live database
• Claude AI writes the narrative sections
• The Hugo static site is rebuilt and deployed to Cloudflare Pages
• Posts are published at approximately 03:00 UTC daily