Trading Strategies Explained

Machine Learning-Based Strategy

What it is: Use algorithms that learn patterns from historical data to predict future price movements.

Example: “RandomForest model learns: When RSI=28 + Volume spike + EMA rising + ATR moderate → 78% chance price goes up in next 25 minutes.”

Momentum Strategy

What it is: Trade in the direction of strong price movement using indicators like RSI or MACD.

Traditional: “If RSI < 30, buy.” — Our ML: Learns when RSI matters and when it doesn’t.

VWAP Trading Strategy

What it is: VWAP (Volume-Weighted Average Price) is the average price weighted by trading volume throughout the day. Traders believe if current price is below VWAP, the stock is “undervalued.”

Traditional: “When price drops 2% below VWAP, buy.” — Our ML: Discovers “price below VWAP + volume spike → 80% UP” vs “price below VWAP + low volume → 50% UP (don’t trade)”

Volatility Trading Strategy

What it is: Trade based on how much prices are moving. Uses Bollinger Bands or ATR (Average True Range).

Our ML advantage: Learns when volatility means opportunity vs. danger. Fixed rules can’t distinguish.

Pattern Recognition

What it is: Identify chart patterns (head and shoulders, triangles, flags) that historically precede price moves.

Limitation: Subjective — what looks like a pattern to one trader looks different to another.

🕯️ Candlestick Patterns

What they are: Candlestick patterns (Hammer, Doji, Engulfing, Shooting Star, Marubozu, Tweezer, etc.) are visual formations in price charts that traders use to predict reversals or continuations. They’re among the most popular topics in trading education.

How they work: Each pattern is a specific relationship between Open, High, Low, and Close within one or two consecutive bars:

Why the system already captures this: Through thousands of training examples, the model can learn:

• “When close is near high AND low is far below open (= Hammer) AND RSI < 35 → 76% chance UP”
• “When open ≈ close (= Doji) AND ATR is high AND volume spikes → trend reversal likely”

ML advantage over traditional candlestick trading:

Traditional approach says: “Hammer at support = ALWAYS buy.” The RandomForest learns from data:

• “Hammer + high volume + RSI oversold → 78% UP” — Trade!
• “Hammer + low volume + downtrend accelerating → 51% UP” — Skip!
• “Hammer works better on AAPL at 10am than at 3:30pm” (time-aware)
• “Engulfing patterns are more reliable after 3+ red bars than after 1” (context-aware)

Scalable by design: The system’s feature engineering pipeline is modular — adding explicit candlestick pattern features (body ratios, wick ratios, pattern flags) is a straightforward extension that further sharpens the model’s pattern recognition without changing the existing architecture.

📊 Bollinger Band Breakout Strategy (Case Study)

What it is: Combines Bollinger Bands with EMA alignment to catch volatility breakouts.

How it works:

1. Setup: Price consolidates inside narrow BBs for 5–8 days (low volatility)
2. EMA Alignment: Moving averages in bullish order (8 > 21 > 34 > 55)
3. Entry: Fast EMA crosses above slower EMA — confirms momentum
4. Exit: Trail for 5–7 candles or price closes below 8 EMA

Three implementation options:

1. Implicit Learning (Current) — ML already sees all components. No action needed.
2. Hybrid Strategy (Future) — ML predictions + BB breakout filter.
3. Pure BB Breakout (Baseline) — Standalone for A/B testing vs ML.

Supertrend Strategy (Case Study)

What it is: A popular trend-following indicator combining ATR with price action. Widely used on TradingView with Pine Script.

How it works:

1. Calculate baseline from (High + Low) / 2
2. Add volatility buffer: ATR × Multiplier (typically 3)
3. Price above Supertrend line → Bullish (green); below → Bearish (red)
4. BUY: When Supertrend flips red → green; SELL: green → red

Scalping Strategy

What it is: Very short-term trading (1–5 minutes) aiming for small, quick profits.

Adaptive Strategy

What it is: Changes behavior based on market conditions (trending vs. choppy, high vs. low volatility).

Sentiment Analysis Trading

What it is: Trade based on news sentiment and social media mood.

Deep Reinforcement Learning

What it is: Advanced ML where an agent learns by trial and error, receiving rewards for profitable trades.

Phase 7+ Start with RandomForest (simpler, proven) before moving to complex deep learning. Requires extensive simulation infrastructure.

Genetic Algorithms

What it is: Optimization technique — automatically test thousands of parameter combinations to find the best settings.

Future Use cases: Optimize tier thresholds (65/70/75%), stop-loss %, position sizing multipliers, confidence thresholds.

These are about HOW to execute orders, not WHEN to trade. They’re execution mechanics.

Mean Reversion ML Learns

Bet that extreme price movements will “revert to the mean.” Uses BB, RSI, moving averages — all already calculated.

Breakout Strategies ML Learns

Trade when price breaks above resistance or below support. Price, BB, volume, and ATR are all available — ML discovers true breakouts vs. false breakouts.

Gap Trading Easy Add

Trade stocks that open significantly higher/lower than previous close. Would need a gap_pct feature added to Stage 2 feature engineering.

News-Based Trading Phase 5

Trade immediately after news releases. The current approach uses news sentiment as one of many features (aggregated over time), not instant reaction — avoiding false signals from market overreaction.

Grid Trading Could Add

Place buy/sell orders at regular price intervals. Requires limit orders and works best in ranging/sideways markets. Different from the directional prediction approach.

Pre-Market / After-Hours Could Extend

The infrastructure supports any hours. Just need data during extended hours and separate ML models (extended hours behave differently).

Arbitrage N/A

Exploit price differences across markets. Requires simultaneous multi-asset positions, direct market access, microsecond execution. Fundamentally different architecture.

High-Frequency Trading N/A

Thousands of trades per second. Requires co-located servers, specialized hardware, multi-million dollar infrastructure. System timeframe: 5 minutes to 5 hours. HFT: sub-second.

Dark Pool Trading N/A

Execute large orders in private exchanges. Institutional only — not accessible to retail traders.

Options Strategies N/A

Trade options contracts (delta neutral, covered calls, iron condors). Different asset class requiring pricing models (Black-Scholes), Greeks. Future phase after stocks prove profitable.

Index Rebalancing N/A

Trade stocks being added/removed from S&P 500. Infrequent quarterly events, institutionally dominated.

Order Flow Imbalance N/A

Trade based on buy/sell order imbalance in the order book. Requires Level 2 market data (expensive, high-frequency). The system uses aggregated OHLCV bars.

Pairs Trading Architecture Mismatch

Trade two related stocks simultaneously (long A, short B). Requires combined P&L tracking and correlation analysis. The system trades symbols independently. Could add but requires significant refactoring.

Seasonality Trading Feature Only

Trade based on calendar patterns (“stocks rise on Fridays”). Too simple as standalone strategy. Could add day_of_week, time_of_day, days_until_earnings as features to Stage 2.

Simple strategies are easy to understand, but markets are complex.

Data shows: Basic moving average crossovers have ~50–55% win rates (barely better than a coin flip). Our ML models achieve 60–70% accuracy on 3-class predictions (UP/DOWN/FLAT).

The problem with simple: When everyone knows a pattern, it stops working. MA crossovers have been public knowledge for 40+ years. Our ML discovers fresh patterns from recent data.

Possibly, but it doesn’t scale and has human limitations.

Best case: Use both! Human oversight with ML execution.

Pairs trading requires simultaneously holding two positions (long A, short B) and tracking combined P&L. Works best in market-neutral hedge funds.

Our approach: Directional trading on individual stocks. We predict if Stock A will go up or down, not if it will outperform Stock B.

Could we add it? Yes, but it requires significant architecture changes. We’d prioritize after validating the current approach.

We are! It’s Phase 5 of the roadmap. News sentiment from professional financial media (Alpha Vantage News API).

Why not social media? Twitter/Reddit sentiment is noisy and easily manipulated. Professional financial news provides higher signal quality.

When useful: Earnings announcements, major news events, market-moving headlines.

The strategy is just one component. The infrastructure, data quality, and continuous improvement create sustainable competitive advantage:

1. Model training process: Retrain 2–3x daily with fresh data
2. Feature engineering: Specific combination of 21 indicators
3. Multi-horizon consensus: How we combine 4 time horizons
4. Infrastructure: Automated data pipelines, error handling, failover
5. Continuous improvement: By the time someone copies v1.0, we’ve moved to v2.0

Risk management and system validation. Paper trading proves:

1. Technical execution works: Orders execute, positions track, errors handled
2. Models perform out-of-sample: Lab accuracy translates to live market accuracy
3. Operational reliability: System runs 10+ hours daily without intervention
4. Financial viability: Positive returns after slippage and fees

Timeline: Move to real money when paper trading demonstrates consistent profitability over 3+ months.

Our advantage: We’re nimble. Banks can’t trade small-cap stocks (not enough liquidity for their size). We can trade anything.