Chapter 3: Building Trading Signals¶

Transform data into actionable trading signals.

What is a Signal?¶

A signal is a numerical score that predicts expected returns:

Higher score → expect higher returns → buy more
Lower score → expect lower returns → sell/short

Text Only

signal momentum:
  // Stocks with higher past returns get higher scores
  emit zscore(ret(prices, 60))

Signal Properties¶

Good Signals Have:¶

Property	Description	Example
Predictive	Forecasts future returns	High momentum → high future returns
Stable	Doesn't change too rapidly	Monthly, not minute-by-minute
Diversified	Works across many assets	Not just one stock
Economically sensible	Has logical explanation	"Winners keep winning"

Signal Quality Metrics¶

Information Coefficient (IC): Correlation between signal and forward returns.

Text Only

IC = 0.05 → Good signal
IC = 0.10 → Excellent signal
IC > 0.15 → Suspicious (check for errors)

Building Your First Signal¶

Step 1: Start with a Hypothesis¶

"Stocks with strong recent performance continue to outperform."

Step 2: Implement Mathematically¶

Text Only

signal momentum:
  // Calculate 60-day return
  past_return = ret(prices, 60)

  // Standardize cross-sectionally
  emit zscore(past_return)

Step 3: Test and Validate¶

Text Only

portfolio test_momentum:
  weights = rank(momentum).long_short(top=0.2, bottom=0.2)
  backtest from 2015-01-01 to 2024-12-31

Common Signal Types¶

Momentum Signals¶

Past returns predict future returns:

Text Only

// Simple momentum
signal momentum_simple:
  emit zscore(ret(prices, 60))

// 12-1 momentum (skip last month)
signal momentum_12_1:
  ret_12m = ret(prices, 252)
  ret_1m = ret(prices, 21)
  emit zscore(ret_12m - ret_1m)

// Relative strength
signal relative_strength:
  stock_ret = ret(prices, 60)
  market_ret = ret(market, 60)
  emit zscore(stock_ret - market_ret)

Value Signals¶

Cheap stocks outperform:

Text Only

// Book-to-market
signal value_btm:
  emit zscore(book_value / market_cap)

// Earnings yield
signal value_ey:
  emit zscore(earnings / prices)

// Combined value
signal value_composite:
  btm = zscore(book_value / market_cap)
  ey = zscore(earnings / prices)
  emit 0.5 * btm + 0.5 * ey

Quality Signals¶

High-quality companies outperform:

Text Only

// Profitability
signal quality_roe:
  emit zscore(net_income / equity)

// Earnings stability
signal quality_stability:
  emit -zscore(rolling_std(earnings, 8))

// Combined quality
signal quality_composite:
  roe = zscore(net_income / equity)
  stability = -zscore(rolling_std(earnings, 8))
  emit 0.6 * roe + 0.4 * stability

Low Volatility Signals¶

Low-risk stocks outperform (risk-adjusted):

Text Only

// Inverse volatility
signal low_vol:
  vol = rolling_std(ret(prices, 1), 60) * sqrt(252)
  emit -zscore(vol)

// Beta
signal low_beta:
  stock_ret = ret(prices, 1)
  market_ret = ret(market, 1)
  beta = rolling_cov(stock_ret, market_ret, 60) / rolling_var(market_ret, 60)
  emit -zscore(beta)

Mean Reversion Signals¶

Extreme moves revert:

Text Only

// Price vs moving average
signal mean_reversion:
  ma = rolling_mean(prices, 20)
  deviation = (prices - ma) / ma
  emit -zscore(deviation)

// RSI-based
signal rsi_reversion:
  change = diff(prices, 1)
  gains = where(change > 0, change, 0)
  losses = where(change < 0, -change, 0)
  rs = ema(gains, 14) / (ema(losses, 14) + 0.0001)
  rsi = 100 - 100 / (1 + rs)
  emit zscore(50 - rsi)

Signal Transformation¶

Standardization¶

Make signals comparable:

Text Only

// Z-score (mean=0, std=1)
signal standardized:
  emit zscore(raw_signal)

// Rank (uniform distribution)
signal ranked:
  emit rank(raw_signal) / count(raw_signal)

Winsorization¶

Handle outliers:

Text Only

signal winsorized:
  lower = quantile(raw_signal, 0.01)
  upper = quantile(raw_signal, 0.99)
  emit clip(raw_signal, lower, upper)

Sector Neutralization¶

Remove sector bias:

Text Only

signal sector_neutral:
  // Demean within each sector
  emit neutralize(raw_signal, by=sectors)

Smoothing¶

Reduce noise:

Text Only

signal smoothed:
  // Exponential moving average of signal
  emit ema(raw_signal, 5)

Combining Signals¶

Simple Average¶

Text Only

signal combined_avg:
  emit (momentum + value + quality) / 3

Weighted Average¶

Text Only

signal combined_weighted:
  emit 0.4 * momentum + 0.4 * value + 0.2 * quality

IC-Weighted¶

Weight by predictive power:

Text Only

signal combined_ic:
  // Weights based on historical IC
  ic_mom = 0.04
  ic_val = 0.03
  ic_qual = 0.02
  total = ic_mom + ic_val + ic_qual

  emit (ic_mom * momentum + ic_val * value + ic_qual * quality) / total

Rank Then Average¶

Text Only

signal combined_rank:
  r_mom = rank(momentum)
  r_val = rank(value)
  r_qual = rank(quality)
  emit (r_mom + r_val + r_qual) / 3

Signal Decay¶

Signals lose predictive power over time:

Text Only

Day 1: IC = 0.05
Day 5: IC = 0.04
Day 21: IC = 0.03
Day 63: IC = 0.01

Managing Decay¶

Text Only

// More frequent rebalancing
portfolio fast_decay:
  weights = ...
  backtest rebal=5 from 2015-01-01 to 2024-12-31  // Every 5 days

// Less frequent for slow decay
portfolio slow_decay:
  weights = ...
  backtest rebal=21 from 2015-01-01 to 2024-12-31  // Monthly

Complete Example¶

Multi-Factor Signal¶

Text Only

data:
  source = "prices_fundamentals.parquet"
  format = parquet

// Individual factors
signal momentum:
  ret_12m = ret(prices, 252)
  ret_1m = ret(prices, 21)
  emit neutralize(zscore(ret_12m - ret_1m), by=sectors)

signal value:
  btm = zscore(book_value / market_cap)
  ey = zscore(earnings / prices)
  raw = 0.5 * btm + 0.5 * ey
  emit neutralize(raw, by=sectors)

signal quality:
  prof = zscore(net_income / equity)
  stab = -zscore(rolling_std(earnings, 8))
  raw = 0.6 * prof + 0.4 * stab
  emit neutralize(raw, by=sectors)

signal low_vol:
  vol = rolling_std(ret(prices, 1), 60) * sqrt(252)
  emit neutralize(-zscore(vol), by=sectors)

// Combined signal
signal multi_factor:
  emit 0.30 * momentum + 0.30 * value + 0.25 * quality + 0.15 * low_vol

// Portfolio
portfolio main:
  weights = rank(multi_factor).long_short(top=0.2, bottom=0.2, cap=0.03)

  constraints:
    max_sector = 0.20

  costs = tc.bps(10)

  backtest from 2015-01-01 to 2024-12-31

Best Practices¶

1. Start Simple¶

Text Only

// Start here
signal simple:
  emit zscore(ret(prices, 60))

// Not here
signal complex:
  emit complicated_formula(many, parameters, here)

2. Use Economic Intuition¶

Ask: "Why should this work?"

3. Normalize Everything¶

Text Only

// Always standardize
emit zscore(raw_signal)

// Or rank
emit rank(raw_signal)

4. Sector Neutralize¶

Text Only

// Avoid unintended sector bets
emit neutralize(signal, by=sectors)

5. Test Out-of-Sample¶

Use walk-forward validation to avoid overfitting.

Exercises¶

Build a momentum signal using 12-1 month returns
Create a value signal combining book-to-market and earnings yield
Combine momentum and value with equal weights
Sector-neutralize your combined signal

Next Chapter¶

Continue to Chapter 4: The sigc Language to master the DSL syntax.