106 terms across quantitative finance, risk management, and machine learning
106 terms
How capital is distributed to a strategy position - fixed dollar amount or a percentage of account equity.
Excess return above what the CAPM model predicts given the strategy's beta to the benchmark.
Total return scaled to a per-year figure, enabling comparison across strategies with different test lengths.
A statistical test for stationarity - whether a time series has a constant mean and variance over time.
Measures market volatility by averaging the true range over N periods - used for dynamic stop loss and take profit placement.
Systematic search over strategy parameters to find the configuration with the best historical performance.
The method used to aggregate price data into OHLCV bars - time, volume, tick, dollar, or ML-optimized.
A trending downward market state with sustained negative momentum and rising risk.
Market sensitivity. β=1 moves in lockstep with benchmark; β=0 uncorrelated; β<0 inversely correlated.
< 0.5 defensive · 0.5–0.9 low · 0.9–1.1 market-like · > 1.1 high · > 1.8 leveraged
A trending upward market state characterized by sustained positive momentum and rising prices.
Compound Annual Growth Rate - the annualized rate at which the portfolio grows if it grew at a constant rate.
Annual return divided by maximum drawdown - measures return per unit of worst-case loss.
> 0.5 acceptable · > 2.0 strong
Up capture divided by down capture. Above 1 = earns more in rising markets than it loses in falling markets.
< 1 unfavorable asymmetry · 1–1.2 slight edge · > 1.2 clear positive asymmetry
A trailing stop that hangs from the highest high like a chandelier - exits when price drops ATR × multiplier from the peak.
The gold-standard time-series CV - generates multiple train/test paths, tests on out-of-sample groups, and computes PBO.
OOS groups: 1–2 for short datasets, 2–3 for long datasets. N paths: auto (combinations) or capped at 20.
A weighted combination of multiple performance metrics used as the optimization objective - balances returns, risk, and stability.
Pearson correlation with benchmark returns. −1 = perfectly inverse, 0 = uncorrelated, +1 = perfectly correlated.
−0.2 to 0.2 uncorrelated · −0.5 to −0.2 weakly inverse · > 0.6 highly correlated
Tests how strategy performance degrades under higher trading costs - reveals whether edge survives realistic execution.
A model validation technique that tests performance across multiple train/test splits to estimate generalization.
Average loss in the worst 5% of periods. Answers "how bad is it when it IS bad?" VaR only gives the threshold.
How historical data is partitioned into training, validation, and holdout sets for model development.
Adjusts Sharpe ratio for the number of trials performed - a high Sharpo after 10,000 trials is less impressive than after 10.
< 1.0 not significant · > 1.0 statistically significant · > 2.0 robust edge
Whether a strategy trades long only, short only, or both directions.
Percentage of benchmark downside the strategy captures. Lower is better - 0% = perfect downside protection.
Standard deviation of returns below zero (or a target). Pure downside risk, used in the Sortino ratio.
How long a drawdown episode lasts from peak to full recovery. Longer durations are harder to endure.
Standard deviation of the drawdown level over time. High values mean losses are erratic and unpredictable.
Combines multiple models into one predictor - reduces variance, improves robustness, and leverages diverse signals.
The time series of portfolio value from start to end of the backtest. The visual fingerprint of a strategy.
Average expected profit or loss per trade, in dollar or percent terms.
How stable the gross exposure level is across all periods. High score = predictable, disciplined position sizing.
Quantifies how much each input feature contributes to model predictions - reveals which signals drive the strategy.
An automated sequence of transformations that clean, filter, and prepare raw features for ML model training.
A method to make a time series stationary while preserving as much memory (predictive signal) as possible.
Average winning trade divided by average losing trade (in absolute terms). Above 1 = wins outsize losses.
|Long| + |Short| - total capital at risk regardless of direction. Unlike net exposure, it never cancels.
The duration a position is held before being closed. Average holding period characterizes the strategy's timeframe.
The final unbiased test - evaluates the selected model on completely unseen data that was never used during optimization.
An automated Bayesian search process that efficiently explores the parameter space using trial history to find optimal configurations.
Correlation between a feature's value today and the forward return - measures predictive power of a single feature.
Active return divided by tracking error. Measures how efficiently the strategy bets against the benchmark.
< 0 active bets destroy value · 0–0.5 neutral · 0.5–1.0 good · > 1.0 excellent
A smooth continuous curve fitted to the return histogram, revealing the underlying probability density.
Mathematically optimal fraction of capital to bet per trade to maximize long-run growth.
10–30% is practical sizing · < 0 = no edge · > 50% = very high risk
Tail heaviness relative to a normal distribution. Values above 3 mean more frequent extreme events.
≈ 3 normal · > 3 fat tails · > 6 very heavy tails (excess kurtosis > 3)
Multiplier on position size relative to account capital. 2× leverage means $1 of capital controls $2 of exposure.
1× no leverage · 2–5× moderate · >10× high (crypto perps)
Isolated: each position has its own margin cap. Cross: all positions share a common margin pool.
A distinct, persistent state of the market (bull, bear, range, transition) identified by a classification model.
The largest peak-to-trough decline in portfolio value. The worst-case loss an investor would have experienced.
< 10% conservative · < 20% acceptable · > 40% high risk
A secondary ML model trained to filter primary predictions - predicts whether a trade signal will be profitable.
Automated hyperparameter search to find the best ML model configuration for a given dataset and horizon.
Generates thousands of synthetic equity curves to stress-test whether observed performance is robust or luck-driven.
Statistical adjustment that prevents false discoveries when testing thousands of strategy configurations.
Long exposure − short exposure. Near zero = market-neutral. Positive = net long bias.
Adds random noise to input data and tests if strategy performance survives - robust strategies degrade gracefully.
Ratio of probability-weighted gains above a threshold to losses below it. Above 1 means gains outweigh losses.
> 1.0 = net gain · > 1.5 good · > 3.0 excellent
Strategy total return minus benchmark total return over the same period.
Multi-objective optimization that finds solutions where improving one metric would worsen another - the Pareto frontier.
Percentage of trading periods where at least one position is open. High coverage = strategy is almost always deployed.
One-way: only one direction per symbol at a time. Hedge: simultaneous long and short positions allowed.
How many bars ahead an ML model is trained to predict. Shorter horizons = faster signals but more noise.
The probability that the true Sharpe ratio is greater than zero, accounting for sample size and non-normality.
> 75% marginal · > 90% likely real edge · > 95% statistically robust
Adjusts raw model probabilities so they reflect true likelihood - a predicted 80% should happen ~80% of the time.
The probability that the selected strategy is overfit - above 0.5 means luck likely explains the backtest more than skill.
< 0.1 robust · 0.1–0.2 acceptable · 0.2–0.5 high risk · > 0.5 probable overfit
Gross profits divided by gross losses. Above 1 = strategy is profitable in aggregate.
> 1.0 profitable · > 1.5 acceptable · > 2.0 strong
An embargo period between training and testing sets to prevent information leakage from overlapping labels.
K-fold cross validation adapted for time series - purges overlapping data between adjacent folds to prevent leakage.
A 0–100 score rating how suitable a bar series is for ML training based on distribution properties.
< 50 poor · 50–70 acceptable · 70–85 good · > 85 excellent
A sideways, consolidating market where price oscillates in a defined band without a clear trend.
Net return divided by max drawdown. Values above 1 mean the strategy has earned back more than its worst loss.
> 1.0 recovered · > 2.0 good · > 5.0 excellent
The model's probability score for the classified regime. Higher confidence = cleaner regime signal.
< 0.5 ambiguous · 0.5–0.7 moderate · > 0.7 high · > 0.85 strong signal
Tests whether strategy performance holds across different market regimes - a strategy that only works in bulls is fragile.
How well a strategy's historically strong regimes match the current market regime. Higher = better alignment.
< 0.3 poor fit · 0.3–0.6 moderate · 0.6–0.8 good · > 0.8 excellent alignment
A shift from one market regime to another. Transitions are periods of elevated uncertainty and model ambiguity.
The full histogram of per-period returns, showing how frequently different return magnitudes occur.
Sharpe ratio computed on a rolling window, showing how risk-adjusted performance evolves over time.
Mathematical probability of losing the entire account given current win rate, avg win/loss, and position sizing.
0% ideal · < 1% acceptable · > 5% dangerous
Per-observation importance weights that adjust how much influence each training sample has on model fitting.
Tests how much strategy performance changes when parameters are perturbed - fragile strategies degrade sharply under small changes.
A bootstrap resampling method for time series that reduces sampling dependence between overlapping observations.
Risk-adjusted return: excess return per unit of total volatility. Higher is better.
> 1.0 acceptable · > 2.0 strong · > 3.0 excellent
The number of actionable trade signals produced per bar. Higher density = more trading opportunities per unit of data.
Asymmetry of the return distribution. Negative = more extreme losses; positive = more extreme gains.
< −1 strong left tail · −1 to 0 mild left · 0 to +1 mild right · > +1 strong right tail
The difference between the expected fill price and the actual execution price. A backtest cost assumption.
0.01–0.05% realistic for BTC/ETH perps at moderate size
Like Sharpe, but only penalizes downside volatility - ignores upside swings.
> 1.5 acceptable · > 2.5 strong
A predefined price level at which a losing trade is automatically closed to limit downside.
Identifies regime-changing discontinuities in the price series where statistical properties shift permanently.
Generates artificial price series with similar statistical properties to test whether the strategy exploits real patterns.
Measures the statistical robustness of a trading system's edge - how consistently it generates returns relative to variance.
1–2 average · 2–3 good · 3–5 excellent · > 5 superb
Right tail size divided by left tail size. Above 1 = heavier right (gain) tail; below 1 = heavier left (loss) tail.
< 0.8 unfavorable · 0.8–1.2 roughly symmetric · > 1.2 favorable right tail
A predefined price level at which a winning trade is automatically closed to lock in gains.
Tests whether optimal parameters are consistent across time periods - unstable parameters suggest overfitting.
Closes a trade automatically after a set duration, regardless of profit or loss - prevents capital from being tied up indefinitely.
The total number of round-trip trades (entry → exit) executed during the backtest or live period.
Volatility of the strategy's active returns (excess over benchmark). Measures how far the strategy drifts from the benchmark.
How often a strategy enters and exits positions - from high-frequency intraday to low-frequency swing.
A stop loss that moves with the price as it trends in your favor, locking in profits dynamically.
Labels each bar as buy/sell/hold based on which of three barriers (profit target, stop loss, time) is hit first.
Combines depth and duration of drawdowns - how long the strategy stays underwater. Lower is always better.
Percentage of benchmark upside the strategy captures. 100% = moves identically in rising markets.
Standard deviation of returns above the mean - pure upside variability.
The loss threshold exceeded only 5% of the time. 95% VaR = you lose more than this only 1 in 20 periods.
Standard deviation of returns scaled to a yearly figure. Measures the typical size of up and down swings.
Coefficient of variation of rolling volatility windows. Low values mean risk is predictable and stable.
Validates strategy performance across multiple rolling time windows - the strongest test of temporal robustness.
Percentage of trades that are profitable. Does not tell you if the strategy is profitable without knowing the win/loss ratio.
Clamps extreme outlier values to a specified percentile to reduce their influence on model training.