Systematic vs. discretionary analysis
Discretionary chart reading relies on subjective pattern recognition where two experienced traders can look at the same chart and reach opposite conclusions. Systematic technical analysis eliminates this ambiguity by encoding every decision into explicit, testable rules. If the 50-day moving average crosses above the 200-day average, the system goes long. There is no room for interpretation, no dependence on mood or cognitive bias, and every signal can be verified against historical data.
Trend following and time-series momentum
Moskowitz, Ooi, and Pedersen (2012) examined time-series momentum across 58 liquid instruments and found that an asset's own past 12-month return significantly predicts its future return, a phenomenon distinct from cross-sectional momentum. Hurst, Ooi, and Pedersen (2017) extended this analysis back to 1880, confirming positive trend-following returns in every decade since, including during major equity drawdowns. This evidence suggests trend following captures a genuine and persistent feature of financial markets rather than a statistical artifact.
Why momentum works
Behavioral finance offers several explanations for momentum's persistence. Initial underreaction occurs when investors are slow to incorporate new information, a consequence of anchoring and conservatism bias (Barberis, Shleifer, and Vishny, 1998). Daniel, Hirshleifer, and Subrahmanyam (1998) attribute trend continuation to overconfidence, while Hong and Stein (1999) propose gradual information diffusion across participants. Because these tendencies are rooted in human psychology, they are unlikely to be fully arbitraged away, explaining why momentum has survived decades of scrutiny.
Adaptive systems and volatility regimes
Fixed-parameter systems assume markets behave the same way in all conditions. A 20-day breakout channel that works in a trending, low-volatility environment generates excessive whipsaws during choppy, high-volatility regimes. Adaptive systems address this by adjusting lookback windows, thresholds, or position sizes based on current conditions. When realized volatility expands, longer lookbacks filter out noise; when volatility contracts, shorter windows capture trend changes faster. This acknowledges that markets shift between regimes, and any system ignoring this will underperform during transitions.
Risk-adjusted performance measurement
Raw returns are misleading without context. A strategy returning 15% annually with 30% maximum drawdown differs fundamentally from one returning 12% with 10% drawdown. The Sharpe ratio normalizes returns by volatility, the Sortino ratio penalizes only downside deviation, and the Calmar ratio divides annualized return by maximum drawdown to measure recovery efficiency. Evaluating strategies across all three metricss whether performance comes from genuine edge or from excessive risk-taking.
Backtesting pitfalls
An impressive backtest is not necessarily a reliable guide to future performance. Overfitting occurs when a model matches historical noise rather than underlying patterns. Look-ahead bias uses information unavailable at trade time, such as revised economic data. Survivorship bias inflates results by testing only on assets that still exist, ignoring delistings. Data snooping arises when researchers test hundreds of parameter combinations and report only the best result without adjusting for the number of trials.
Out-of-sample testing and walk-forward validation
The most effective safeguard against overfitting is strict separation between in-sample and out-of-sample data. Walk-forward validation extends this by repeatedly training on a rolling window and testing on the subsequent period, producing genuine out-of-sample results. Combinatorial purged cross-validation (CPCV), introduced by de Prado (2018), generates multiple train-test splits while purging overlapping data to prevent information leakage. These methods do not guarantee future performance but substantially reduce the probability of deploying strategies built on statistical illusions.
