Overview
Multicollinearity occurs when two or more independent (predictor) variables in a regression are highly linearly related. It does not bias the estimated coefficients, but it inflates their variances, making coefficient estimates unstable and hypothesis tests unreliable. In investment analysis this commonly appears when multiple technical or fundamental indicators are derived from the same underlying price or accounting series.
Key takeaways
– Multicollinearity makes it difficult to interpret the contribution of individual predictors because standard errors grow and t‑statistics shrink.
– Perfect multicollinearity (exact linear dependence) prevents estimation altogether; high multicollinearity produces very imprecise estimates.
– Detection tools include correlation matrices, Variance Inflation Factor (VIF), tolerance, condition indices, and variance‑decomposition proportions.
– Remedies include dropping or combining predictors, collecting new data, transforming variables, or using regularized / dimension‑reduction regressions (ridge, LASSO, principal component regression, partial least squares).
– In trading/investment work: avoid redundant indicators; combine different indicator families (momentum, trend, volatility, volume) rather than many variants of the same family.
Understanding the basics
– Regression context: You predict a dependent variable Y (e.g., future stock returns) using multiple Xs (e.g., P/E ratio, market cap, momentum).
– Multicollinearity arises when some Xs are correlated with each other (not truly independent).
– Example: Past performance and market capitalization might correlate because larger firms historically deliver different returns; using both without checking correlation may create multicollinearity.
Types of multicollinearity
1. Perfect multicollinearity
• An exact linear relationship among predictors (e.g., X3 = 2*X1 + 3*X2). The design matrix is singular and OLS cannot be computed.
2. High (but not perfect) multicollinearity
• Predictors are highly—but not exactly—linearly related. OLS works but coefficients are imprecise.
3. Structural multicollinearity
• Caused by how variables are constructed (e.g., deriving many indicators from the same price series). Common in technical analysis.
4. Data‑based multicollinearity
• Caused by the data collection process (e.g., small samples, lack of variation, observational data with co-movement).
Why multicollinearity is a problem
– Inflated standard errors → wider confidence intervals → fewer statistically significant coefficients.
– Unstable coefficient estimates: small changes to the sample can flip signs or magnitudes.
– Difficulty in interpreting marginal effects: you cannot reliably say how much Y changes holding the other Xs fixed.
– In extreme cases (perfect multicollinearity) the regression cannot be estimated at all.
How to detect multicollinearity
1. Pairwise correlation matrix
• Quick look: large absolute correlations (e.g., |r| > 0.8) can indicate trouble, though multicollinearity can exist without very high pairwise correlations.
2. Variance Inflation Factor (VIF)
• Formula: VIF_j = 1 / (1 − R_j^2), where R_j^2 is the R² from regressing predictor j on the other predictors.
• Rules of thumb: VIF = 1 (no correlation); 1 < VIF 5 or > 10 (high multicollinearity). Investopedia cites 1–5 moderate, 5–10 high.
• Tolerance = 1 / VIF.
3. Condition index and eigenvalues (Collinearity diagnostics)
• Condition number (square root of max eigenvalue / min eigenvalue) > 30 often indicates multicollinearity; > 100 is severe.
4. Variance‑decomposition proportions
• Show which coefficients’ variances are impacted and may help isolate groups of collinear variables.
5. Signs to watch for in output
• Large standard errors, small t‑stats, coefficients with counterintuitive signs that change a lot with variable selection.
Practical detection commands (examples)
– R: car::vif(lm_model)
– Python (statsmodels):
from statsmodels.stats.outliers_influence import variance_inflation_factor
VIF = [variance_inflation_factor(X.values, i) for i in range(X.shape[1])]
– Also compute pairwise correlations with pandas .corr() or R cor().
Practical consequences for investment analysis
– Using multiple technical indicators derived from the same prices (e.g., Stochastics, Williams %R, RSI) often produces collinear predictors with little independent information.
– Models with collinear fundamental ratios (e.g., different profitability margins that move together) can produce misleading inference about which ratios matter.
– Multicollinearity reduces your ability to rank or test predictors — it’s harder to identify the truly predictive signals.
Effective solutions and practical steps
1. Start with detection
• Step 1: compute pairwise correlations and VIFs for all predictors.
• Step 2: examine condition indices / variance decomposition if you suspect complicated multicollinearity.
2. Decide on action with business/statistical context in mind
• If predictors are redundant for your decision: drop one or more collinear variables.
• If all are important theoretically: combine them into a composite index (e.g., principal component, simple weighted average).
3. Variable transformation and feature engineering
• Centering or standardizing helps with interpretation and sometimes with polynomial/interaction multicollinearity (centering reduces correlation between level and squared terms).
• Use differences, ratios, or orthogonalized residuals (partial out the effect of a strongly collinear predictor).
4. Collect more data
• Better cross‑sectional or time‑varying data can reduce sampling correlation or reveal independent variation.
5. Use alternative modeling techniques
• Ridge regression: shrinks coefficients, reduces variance problems (good for prediction, not for unbiased interpretation).
• LASSO: can select a subset of predictors (helps interpretation and parsimony).
• Principal Component Regression (PCR): regress on orthogonal components (dimension reduction).
• Partial Least Squares (PLS): components that maximize covariance with Y (useful when predictors are many and collinear).
6. For trading/technical analysis specifically
• Use indicators from different families: momentum + volatility + volume + trend.
• Avoid stacking many indicators calculated from the same raw series or with very similar formulas.
• If you must include similar indicators, test for multicollinearity and consider combining them into a single signal.
Trade‑offs and caveats
– Dropping variables reduces multicollinearity but can introduce omitted variable bias if the dropped variable had an independent effect.
– Regularization (ridge, LASSO) improves prediction and numerical stability but changes coefficient interpretation (biased estimates).
– PCA/PCR change variables to components that may be harder to interpret economically.
Interpreting multicollinearity results
– High VIFs tell you the variance of a coefficient is inflated by roughly the VIF factor relative to orthogonal predictors.
– Example: if VIF for X1 = 9, the variance of its estimated coefficient is 9 times what it would be if X1 were uncorrelated with the other Xs.
– Beware: a predictor can have a low VIF but still be involved in multicollinearity through combinations of variables — hence condition diagnostics are useful.
Practical checklist for an analyst/investor (step‑by‑step)
1. Before modeling: list candidate predictors and note which are likely related (e.g., many momentum indicators).
2. Pre‑screen: compute pairwise correlations and a VIF table.
3. If VIFs are high:
• Group similar indicators and keep one representative OR
• Create a composite signal (PCA) OR
• Apply ridge/LASSO for prediction tasks.
4. Re‑estimate model and check coefficient stability (bootstrap or re‑sample to see sensitivity).
5. For inference, favor parsimonious models with interpretable, less collinear predictors.
6. For pure prediction, prefer regularization or dimension reduction and evaluate predictive performance out of sample.
Fast facts
– VIF = 1: no multicollinearity. VIF > 5 (or 10) is usually considered problematic.
– Perfect multicollinearity makes the design matrix singular; you must remove or reparameterize variables.
– Structural multicollinearity is especially common in technical analysis because many indicators derive from the same price/volume series.
Bottom line
Multicollinearity is a practical and common problem in regression-based investment analysis. Detect it early (correlations, VIF, condition indices), then decide whether to remove, combine, transform, or regularize predictors depending on whether your priority is interpretation or prediction. For technical analysis, avoid using many indicators of the same type; combine indicators across different families to preserve independent information.
Further reading / reference
– Investopedia: Multicollinearity — (source for this article)
– For implementation details: statsmodels and scikit-learn documentation (VIF, ridge, lasso, PCA)
Editor’s note: The following topics are reserved for upcoming updates and will be expanded with detailed examples and datasets.