Machine Learning For Portfolio Optimization

Expert-defined terms from the Professional Certificate in Risk Modeling with Machine Learning course at LearnUNI. Free to read, free to share, paired with a professional course.

Machine Learning For Portfolio Optimization

Alpha – Expected excess return of a portfolio relative to a benchmark #

Related: Beta, Sharpe ratio, CAPM. In portfolio optimization, alpha is the objective to maximize; models estimate it from factor exposures. Challenge: Estimating stable alpha in noisy data.

Alpha Decay – The reduction of predictive power of a model over time #

Related: Concept drift, model retraining. Machine‑learning models may capture transient patterns that fade, requiring periodic recalibration to maintain performance.

Amplitude Modulation – Technique to encode information by varying signal… #

Related: Time‑series preprocessing, feature engineering. Rarely used directly in finance but can inspire methods for handling heteroscedasticity.

ARIMA (AutoRegressive Integrated Moving Average) – Classical time‑series… #

Related: SARIMA, Box‑Jenkins methodology. Serves as a baseline for forecasting asset returns before applying more complex ML models.

Backtesting – Simulated evaluation of a strategy on historical data #

Related: Walk‑forward analysis, out‑of‑sample testing. Essential to assess portfolio optimization algorithms; must avoid look‑ahead bias and overfitting.

Bayesian Optimization – Sequential model‑based method for hyperparameter… #

Related: Gaussian processes, acquisition function. Used to select regularization strength or number of trees in ensemble models for portfolio construction.

Beta – Sensitivity of a portfolio’s returns to market movements #

Related: Alpha, systematic risk, CAPM. In risk‑aware optimization, beta constraints limit exposure to market volatility.

Black‑Litterman Model – Framework combining equilibrium market returns wi… #

Related: Reverse optimization, prior distribution. Machine‑learning can generate the view vector from predictive models, improving robustness.

Boosting – Ensemble technique that sequentially adds weak learners to cor… #

Related: AdaBoost, Gradient Boosting, XGBoost. Popular for predicting asset returns; must guard against overfitting due to high variance in financial data.

Box‑Cox Transformation – Power transformation to stabilize variance #

Related: Yeo‑Johnson, log transformation. Applied to skewed return series before feeding into linear or tree‑based models.

Brownian Motion – Continuous‑time stochastic process with independent Gau… #

Related: Geometric Brownian motion, Wiener process. Underlies many asset‑price models and provides simulated scenarios for Monte‑Carlo portfolio evaluation.

Concept Drift – Change in underlying data distribution over time #

Related: Non‑stationarity, adaptive learning. In portfolio optimization, drift may render a once‑effective model obsolete, necessitating online learning or periodic retraining.

Cross‑Validation – Technique to assess model performance on unseen data b… #

Related: K‑fold, time‑series split, rolling window. For financial data, use a forward‑looking split to preserve temporal order.

CVaR (Conditional Value at Risk) – Expected loss beyond the VaR threshold #

Related: VaR, tail risk, coherent risk measures. Optimization often minimizes CVaR to control extreme downside risk.

Data Snooping – Inadvertent use of information from the test set during m… #

Related: Look‑ahead bias, overfitting. Leads to inflated performance estimates; rigorous separation of training and evaluation periods mitigates this risk.

Dimensionality Reduction – Process of reducing the number of variables wh… #

Related: PCA, t‑SNE, autoencoders. Helps alleviate the curse of dimensionality in large asset universes and improves model stability.

Dropout – Regularization technique that randomly omits neurons during tra… #

Related: L1/L2 regularization, early stopping. Prevents overfitting in deep neural networks used for return prediction.

Ensemble Methods – Combining multiple models to improve predictive perfor… #

Related: Bagging, stacking, voting. In portfolio optimization, ensembles often yield more robust forecasts than single models.

Feature Engineering – Creation of informative variables from raw data #

Related: Lagged returns, volatility measures, macro indicators. Critical for ML models; poorly engineered features can degrade performance dramatically.

Feature Selection – Process of choosing a subset of relevant features #

Related: LASSO, recursive feature elimination, mutual information. Reduces overfitting risk and computational load, especially important when the asset universe is large.

Fisher Information – Measure of the amount of information a random variab… #

Related: Cramér‑Rao bound, information matrix. Appears in portfolio theory when assessing estimator efficiency.

GBM (Geometric Brownian Motion) – Model for asset price dynamics assuming… #

Related: Black‑Scholes, Monte‑Carlo simulation. Provides baseline scenario generation for stress testing optimized portfolios.

Gaussian Process – Non‑parametric Bayesian model defining a distribution… #

Related: Kernel methods, Kriging. Useful for quantifying predictive uncertainty in return forecasts, enabling risk‑aware allocation.

Gradient Boosting – Boosting variant that fits residuals using gradient d… #

Related: XGBoost, LightGBM, CatBoost. Dominant in return prediction competitions; hyperparameter tuning crucial to avoid overfitting to noisy financial signals.

Information Ratio – Ratio of portfolio alpha to tracking error #

Related: Sharpe ratio, Sortino ratio. Optimization may target maximization of the information ratio, balancing excess return against active risk.

Kernel Methods – Algorithms that operate in transformed feature spaces #

Related: Support vector machines, Gaussian kernels. Enable nonlinear separation of asset characteristics without explicit feature mapping.

KNN (K‑Nearest Neighbors) – Instance‑based learning that predicts based o… #

Related: Distance metrics, weighting schemes. Simple baseline for return forecasting; suffers from curse of dimensionality in large asset sets.

Lagged Features – Past values of a variable used as predictors #

Related: Autoregressive terms, rolling windows. Common in time‑series models to capture momentum or mean‑reversion effects.

LSTM (Long Short‑Term Memory) – Recurrent neural network architecture des… #

Related: GRU, RNN. Effective at modeling temporal dependencies in price series, yet requires careful regularization to prevent overfitting.

Markowitz Mean‑Variance Optimization – Classical framework balancing expe… #

Related: Efficient frontier, quadratic programming. ML enhancements replace the estimation of means and covariances with data‑driven forecasts.

Maximum Drawdown – Largest peak‑to‑trough decline over a period #

Related: Drawdown risk, recovery time. Often incorporated as a constraint in portfolio construction to limit tail risk.

Monte‑Carlo Simulation – Repeated random sampling to assess portfolio out… #

Related: Scenario analysis, stochastic modeling. Generates distributions of returns based on ML‑predicted parameters, aiding risk assessment.

Momentum – Tendency of assets that performed well to continue doing so in… #

Related: Trend following, mean reversion. Feature engineering frequently includes momentum indicators such as 12‑month cumulative returns.

Multicollinearity – High correlation among predictor variables #

Related: Variance inflation factor, ridge regression. Can inflate coefficient variance in linear models; regularization or dimensionality reduction mitigates the issue.

Neural Architecture Search – Automated process of discovering optimal net… #

Related: Hyperparameter optimization, AutoML. Can produce bespoke models for specific asset classes, though computationally intensive.

Non‑Stationarity – Property of a time series whose statistical characteri… #

Related: Unit root, differencing. Requires techniques like rolling windows or adaptive learning to maintain model relevance.

Overfitting – Model captures noise rather than underlying pattern #

Related: Regularization, cross‑validation. In finance, overfitted models often produce spectacular backtest results but fail in live trading.

Parsimony – Preference for simpler models that achieve comparable perform… #

Related: Occam’s razor, model complexity. Encouraged to improve interpretability and reduce overfitting risk in portfolio optimization.

Passive Investing – Strategy that tracks a benchmark rather than actively… #

Related: Index funds, tracking error. ML can be used to construct smart‑beta indices that blend passive exposure with factor tilts.

Performance Attribution – Decomposition of portfolio returns into source… #

Related: Factor attribution, sector attribution. Helps evaluate whether ML‑driven signals are adding value beyond market movements.

Portfolio Turnover – Frequency of trading required to maintain target wei… #

Related: Transaction costs, liquidity. Optimization often penalizes turnover to balance expected gains against trading expenses.

Predictive Modeling – Building statistical or ML models to forecast futur… #

Related: Regression, classification, time‑series forecasting. Core of ML‑based portfolio construction; quality of predictions drives allocation decisions.

Probabilistic Forecasting – Producing a full distribution rather than a p… #

Related: Quantile regression, Bayesian inference. Enables risk‑aware optimization by incorporating uncertainty directly into the objective.

Quadratic Programming – Optimization method for problems with a quadratic… #

Related: Convex optimization, interior‑point methods. Underpins many mean‑variance formulations; scalable solvers are critical for large asset universes.

Random Forest – Ensemble of decision trees trained on bootstrapped sample… #

Related: Bagging, feature importance. Robust to overfitting and provides interpretable variable importance metrics for return predictors.

Rebalancing Frequency – Interval at which portfolio weights are adjusted #

Related: Daily, monthly, quarterly. Trade‑off between capturing new signals and incurring transaction costs; ML can recommend optimal frequency based on market regime.

Regularization – Adding penalty terms to loss functions to discourage com… #

Related: L1 (LASSO), L2 (Ridge), Elastic Net. Crucial for stabilizing coefficient estimates when training data is limited relative to predictors.

Risk Parity – Allocation method that equalizes risk contribution across a… #

Related: Volatility weighting, diversification. ML can estimate forward‑looking risk measures to improve risk‑parity allocations.

Rolling Window – Moving time window used for model training and evaluatio… #

Related: Expanding window, walk‑forward analysis. Provides up‑to‑date parameter estimates while preserving temporal order.

Scenario Analysis – Examination of portfolio performance under predefined… #

Related: Stress testing, macro shocks. ML‑generated scenarios can reflect realistic joint movements of assets.

Sharpe Ratio – Excess return per unit of total risk (standard deviation) #

Related: Information ratio, Sortino ratio. Often used as an objective function; however, it assumes normally distributed returns.

Shrinkage Estimator – Technique that pulls sample covariance matrix towar… #

Related: Ledoit‑Wolf, Bayesian shrinkage. Improves stability of covariance estimates used in mean‑variance optimization.

Signal‑to‑Noise Ratio – Measure of predictive strength relative to random… #

Related: Information coefficient, t‑statistic. Low ratios in finance demand robust modeling and careful validation.

Simplified Portfolio Theory (SPT) – Approximation that treats assets as i… #

Related: Factor models, diagonal covariance. ML can estimate factor sensitivities to relax independence assumptions.

Smoothing – Technique to reduce volatility in time‑series signals #

Related: Moving average, exponential smoothing. Helps stabilize input features for ML models but may lag behind rapid market changes.

Stochastic Gradient Descent (SGD) – Iterative optimization algorithm usin… #

Related: Mini‑batch, learning rate. Preferred for training large neural networks on high‑frequency financial data.

Stress Testing – Evaluation of portfolio resilience under extreme but pla… #

Related: Scenario analysis, tail risk. ML can generate stress scenarios by perturbing model inputs in realistic ways.

Structural Break – Point at which the statistical properties of a series… #

Related: Regime shift, Chow test. Detecting breaks is vital for updating ML models to maintain predictive accuracy.

Supervised Learning – Learning paradigm where models are trained on label… #

Related: Regression, classification. Most return‑prediction tasks fall under supervised learning, requiring historical return labels.

Support Vector Machine (SVM) – Classification/regression algorithm that m… #

Related: Kernel trick, soft margin. Can be applied to predict binary outcomes such as up/down movements; careful tuning needed for financial noise.

Swaption – Option granting the right to enter an interest‑rate swap #

Related: Derivatives, volatility surface. Pricing models may incorporate ML‑estimated volatility surfaces to improve hedging decisions.

Technical Indicator – Quantitative measure derived from price and volume… #

Related: RSI, MACD, Bollinger Bands. Frequently used as features; must be validated for predictive relevance.

Temporal Fusion Transformer (TFT) – Attention‑based architecture for mult… #

Related: Transformers, variable selection networks. Demonstrated strong performance on financial series where multiple covariates evolve over time.

Transaction Cost Model – Quantitative representation of fees, slippage, a… #

Related: Linear cost, non‑linear impact. Incorporating realistic cost models into optimization prevents over‑trading driven by ML signals.

Training‑Test Split – Division of data into subsets for model fitting and… #

Related: Hold‑out, validation set. In finance, splits must respect chronological order to avoid contaminating future information.

Transfer Learning – Reusing knowledge from a source domain to improve lea… #

Related: Fine‑tuning, domain adaptation. Pre‑trained models on large market data can be adapted to niche asset classes.

Tree‑Based Models – Algorithms that partition feature space using decisio… #

Related: CART, random forest, gradient boosting. Offer interpretability via feature importance and handle mixed data types well.

Trend Following – Strategy that invests in assets exhibiting persistent p… #

Related: Momentum, breakout. ML can enhance trend detection by combining multiple time‑scale signals.

Unsupervised Learning – Learning from data without explicit labels #

Related: Clustering, dimensionality reduction. Useful for discovering latent structures such as hidden market regimes.

Variance Inflation Factor (VIF) – Metric quantifying multicollinearity am… #

Related: Condition number, ridge regression. High VIF values suggest the need for feature elimination or regularization.

Volatility Clustering – Phenomenon where high‑volatility periods tend to… #

Related: GARCH, stochastic volatility. Modeling this effect improves risk forecasts used in portfolio construction.

Weighted Least Squares (WLS) – Regression technique assigning different w… #

Related: Heteroscedasticity, robust regression. Can prioritize recent data points when estimating factor returns.

Yield Curve – Graph of interest rates across different maturities #

Related: Term structure, Nelson‑Siegel model. ML can capture dynamics of the curve for fixed‑income portfolio optimization.

Z‑Score Normalization – Centering data to zero mean and unit variance #

Related: Standardization, scaling. Common preprocessing step for models sensitive to feature magnitude, such as SVMs and neural networks.

Zero‑Mean Portfolio – Portfolio constructed to have zero expected return,… #

Related: Market neutral, hedged. ML can identify assets that offset each other’s exposures, achieving neutrality.

Alpha‑Beta Decomposition – Separation of portfolio return into market‑rel… #

Related: Factor model, residual return. Useful for evaluating the contribution of ML‑generated signals beyond systematic exposure.

Backpropagation – Algorithm for computing gradients in neural networks #

Related: Chain rule, learning rate. Core of training deep learning models for financial forecasting.

Bayesian Inference – Statistical method that updates prior beliefs with o… #

Related: Posterior distribution, conjugate priors. Provides a natural framework for incorporating uncertainty in model parameters used for portfolio allocation.

Bootstrap Resampling – Generating multiple datasets by sampling with repl… #

Related: Confidence intervals, out‑of‑bag error. Enables estimation of variability in performance metrics for ML models in finance.

Capital Asset Pricing Model (CAPM) – Linear model linking expected return… #

Related: Market portfolio, risk‑free rate. Serves as a benchmark for evaluating the added value of ML‑driven alpha.

CatBoost – Gradient‑boosting library that handles categorical features na… #

Related: LightGBM, XGBoost. Often yields superior performance on tabular financial data with mixed variable types.

Change Point Detection – Identifying times where statistical properties s… #

Related: Bayesian online change point detection, CUSUM. Alerts practitioners to moments when model retraining may be required.

Conditional Expectation – Expected value of a variable given certain info… #

Related: Law of iterated expectations, regression. Central to mean‑variance optimization where future returns are conditioned on predictive signals.

Correlation Matrix – Square matrix showing pairwise correlations among as… #

Related: Covariance matrix, eigenvalue decomposition. Accurate estimation is vital for risk budgeting; shrinkage techniques improve stability.

Cross‑Asset Modeling – Simultaneous modeling of multiple asset classes #

Related: Multi‑task learning, hierarchical models. Allows ML to capture inter‑market dynamics that improve diversification benefits.

Decile Portfolio – Grouping assets into ten equal‑size buckets based on a… #

Related: Quantile sorting, factor portfolios. Facilitates evaluation of predictive signals by comparing performance across deciles.

Deep Reinforcement Learning – Combination of deep learning with reinforce… #

Related: Q‑learning, policy gradient. Emerging approach for dynamic portfolio rebalancing under transaction cost constraints.

Diffusion Model – Stochastic process describing how a variable evolves co… #

Related: Ornstein‑Uhlenbeck, mean‑reversion. Provides analytical forms for asset dynamics used in scenario generation.

Elastic Net – Regularization method blending L1 and L2 penalties #

Related: LASSO, Ridge. Useful for feature selection when predictors are highly correlated, as common in financial factor libraries.

Ensemble Kalman Filter – Recursive algorithm for estimating hidden states… #

Related: Data assimilation, particle filter. Can be employed to update latent factor estimates as new market data arrive.

Factor Model – Representation of asset returns as a linear combination of… #

Related: Fama‑French, Barra. ML can estimate factor loadings and residual variances more flexibly than traditional OLS.

Gaussian Mixture Model (GMM) – Probabilistic model assuming data are gene… #

Related: EM algorithm, clustering. Captures multimodal return distributions for scenario analysis.

Hierarchical Risk Parity (HRP) – Portfolio construction method that uses… #

Related: Tree‑based allocation, diversification. Avoids matrix inversion, making it robust to estimation error; ML can refine clustering inputs.

Information Coefficient (IC) – Correlation between predicted and realized… #

Related: Rank IC, predictive power. Serves as a performance metric for ML models; low IC values indicate limited usefulness.

Kernel Density Estimation (KDE) – Non‑parametric method to estimate proba… #

Related: Bandwidth selection, smoothing. Useful for visualizing return distributions and for constructing empirical risk measures.

Knock‑out Options – Derivatives that become worthless if the underlying a… #

Related: Barrier options, exotic derivatives. Pricing models may incorporate ML‑estimated volatility surfaces for more accurate valuation.

Laguerre Polynomials – Orthogonal polynomials used in function approximat… #

Related: Basis expansion, spectral methods. Occasionally employed to model term structure dynamics in fixed‑income portfolio optimization.

Learning Rate – Hyperparameter controlling step size in gradient‑based op… #

Related: Decay schedule, Adam optimizer. Too high a learning rate can cause divergence; too low slows convergence, especially in noisy financial data.

Liquidity Risk – Risk arising from the inability to trade assets without… #

Related: Market impact, bid‑ask spread. Optimization models often include liquidity constraints to avoid excessive exposure to illiquid securities.

Markov Chain Monte Carlo (MCMC) – Class of algorithms for sampling from c… #

Related: Metropolis‑Hastings, Gibbs sampling. Enables Bayesian estimation of portfolio weights when analytical solutions are unavailable.

Mean Reversion – Tendency of asset prices to return to a long‑term averag… #

Related: Ornstein‑Uhlenbeck process, pair trading. ML models may capture mean‑reverting signals through lagged features and regime detection.

Monte‑Carlo Dropout – Technique that interprets dropout at inference time… #

Related: Uncertainty quantification, predictive intervals. Provides a cheap way to assess confidence in neural‑network forecasts for portfolio risk.

Neural Tangent Kernel (NTK) – Analytical tool describing the behavior of… #

Related: Kernel methods, convergence. Offers insight into why certain deep architectures generalize well on financial data.

Optimal Stopping – Decision problem of choosing the time to take a partic… #

Related: American option pricing, dynamic programming. Reinforcement‑learning agents use this framework for trade execution.

Out‑of‑Sample Performance – Measure of model accuracy on data not used du… #

Related: Backtesting, validation set. Critical for assessing whether ML‑driven portfolio strategies will succeed in live markets.

Partial Least Squares (PLS) – Dimensionality reduction technique that pro… #

Related: Canonical correlation analysis, regression. Handles collinear features common in macro‑economic datasets used for asset allocation.

Portfolio Optimization – Process of selecting asset weights to achieve a… #

Related: Mean‑variance, robust optimization. Machine learning enhances the estimation of expected returns and covariances, the core inputs to optimization.

Quantile Regression – Regression technique estimating conditional quantil… #

Related: Expectile regression, VaR estimation. Provides asymmetric loss functions useful for tail‑risk‑focused portfolio construction.

Recursive Feature Elimination (RFE) – Wrapper method that iteratively rem… #

Related: Feature importance, model selection. Helps identify a parsimonious set of predictors for return forecasting.

Regularized Portfolio Optimization – Incorporating penalty terms (e #

G., L1 for sparsity) into the objective function. Related: Elastic net, constrained optimization. Produces more stable and interpretable allocations, especially when the number of assets exceeds observations.

Risk Budgeting – Allocating risk contributions rather than capital #

Related: Risk parity, volatility targeting. ML forecasts of future volatilities feed directly into risk‑budget calculations.

Robust Optimization – Optimization approach that accounts for uncertainty… #

Related: Ambiguity set, worst‑case scenario. Addresses estimation error in ML‑predicted means and covariances, producing portfolios less sensitive to mis‑specification.

Rolling Forecast Origin – Strategy where the start point of the training… #

Related: Expanding window, walk‑forward validation. Mimics real‑world deployment of models that update continuously.

Sample Covariance Matrix – Empirical estimator of asset return covariance… #

Related: Shrinkage estimator, factor model. Often ill‑conditioned in high‑dimensional settings; regularization improves its use in optimization.

Scenario Generation – Creation of plausible future paths for risk factors #

Related: Monte‑Carlo simulation, copula models. ML techniques such as variational autoencoders can produce realistic joint distributions.

Sharpe Ratio Maximization – Objective of maximizing excess return per uni… #

Related: Mean‑variance, utility function. Directly optimizing this ratio is non‑convex; approximations or surrogate objectives are employed.

Shrinkage Target – Structured matrix toward which the sample covariance i… #

Related: Identity matrix, constant correlation. Choosing an appropriate target reduces estimation error in high‑dimensional portfolios.

Smoothing Parameter – Controls the degree of regularization in non‑parame… #

Related: Bandwidth, penalty term. Important when fitting splines to noisy financial data.

Spurious Correlation – Apparent relationship caused by chance rather than… #

Related: Multiple testing, data mining bias. ML pipelines must include statistical controls to avoid exploiting such false patterns.

Stochastic Dominance – Preference ordering of distributions based on expe… #

Related: First‑order, second‑order dominance. Can be incorporated into portfolio selection to ensure dominance over a benchmark.

Structure‑Preserving Transformation – Mapping that maintains relationship… #

G., Monotonicity) between variables. Related: Rank transformation, copula. Useful when normalizing returns while retaining dependence structure for risk modeling.

Super‑hedging – Constructing a portfolio that dominates the payoff of a d… #

Related: Convex risk measures, dual representation. ML can approximate super‑hedging costs by learning worst‑case scenarios.

Synthetic Asset – Constructed security whose payoff mimics a desired expo… #

Related: Replicating portfolio, factor mimicry. ML can design synthetic assets that capture complex factor tilts not directly investable.

Target Volatility Strategy – Portfolio that adjusts exposure to maintain… #

Related: Risk budgeting, volatility scaling. Forecasted volatilities from ML models drive the scaling factor.

Temporal Difference Learning – Reinforcement‑learning method that updates… #

Related: TD(λ), Q‑learning. Applied to learn optimal trading policies that adapt to market dynamics.

Time‑Series Cross‑Validation – Validation technique that respects tempora… #

Related: Rolling origin, forward chaining. Prevents leakage of future information into model training.

Transaction Cost Analysis (TCA) – Assessment of the cost incurred when ex… #

Related: Implementation shortfall, slippage. Incorporating TCA into the optimization objective prevents eroding ML‑generated alpha.

Transformers – Attention‑based neural architectures that process sequence… #

Related: Self‑attention, positional encoding. Emerging models for financial time series, offering improved long‑range dependency capture over RNNs.

Unbiased Estimator – Statistic whose expected value equals the true param… #

Related: Bias‑variance trade‑off, consistency. In finance, unbiasedness may be sacrificed for lower variance via regularization.

Value at Risk (VaR) – Quantile of the loss distribution over a specified… #

Related: CVaR, tail risk. Often used as a constraint in portfolio optimization to limit potential losses.

Variance‑Covariance Matrix – Matrix containing variances along the diagon… #

Related: Correlation matrix, factor decomposition. Accurate estimation is crucial for any risk‑based allocation.

Weighted Portfolio – Portfolio where each asset carries a specific weight… #

Related: Capital allocation, fractional shares. Optimization determines the set of weights that best achieve the objective.

Yield Curve Modeling – Statistical representation of interest rates acros… #

Related: Nelson‑Siegel, spline models. Machine‑learning can capture nonlinear dynamics and improve fixed‑income portfolio decisions.

Z‑Score Portfolio – Portfolio constructed by ranking assets based on thei… #

Related: Mean‑variance, ranking strategy. Allows comparison across assets with differing scales, facilitating cross‑asset allocation.

Zero‑Sum Game – Situation where one participant’s gain equals another’s l… #

Related: Competitive markets, game theory. In portfolio optimization, market impact can be modeled as a zero‑sum interaction among traders.

June 2026 intake · open enrolment
from £90 GBP
Enrol