Glossary
Glossary
Terms used in AI quant, defined plainly. Each entry says what it is and how to check for it.
Fundamentals16
Alpha
Alpha is the return of a strategy above what the market itself delivered. It is the thing everyone claims and nobody can measure from a screenshot — because it decays the moment it becomes known.
Benchmark
A benchmark is the reference a strategy is measured against. Without one, a return number says nothing — because the market itself may have delivered it.
Beta
Beta measures how much a strategy moves with the market. It is the part of a return that comes from market exposure rather than skill — which is why an alpha claim means nothing without it.
Correlation
Correlation measures how much two things move together. In a portfolio it is the difference between diversification and a bigger version of one bet.
Drawdown
Drawdown is the fall from a peak to a later trough, usually expressed as a percentage. Maximum drawdown is the worst such fall — and it is the number that tells you whether you could actually hold a strategy.
Factor
A factor is a measurable characteristic used to explain or predict returns — value, momentum, size, quality. Factors are the raw material of quantitative screening, and the area where data quality matters most.
Kelly criterion
The Kelly criterion sizes a bet from the estimated edge and the odds, to maximise long-run growth. It is elegant, and it is fragile — because it assumes you know the edge exactly.
Liquidity
Liquidity is how easily a position can be traded without moving the price. It is the hidden constraint behind slippage, capacity, and why a backtest that worked at small size may fail at large.
Mean reversion
Mean reversion is the tendency of prices to move back toward an average after a move away from it. It is the mirror of momentum — and the two often operate on different horizons at once.
Momentum
Momentum is the tendency of recent winners to keep winning over some horizon. It is one of the most durable factors — and one of the easiest to overfit, because the horizon is a free parameter.
Position sizing
Position sizing decides how much capital goes into each trade. It often matters more to the outcome than the entry rule — and it is the part most claims skip.
Rebalancing
Rebalancing is restoring a portfolio to its target weights. It sounds like housekeeping, but it is a series of trades — and therefore a cost, and a decision.
Regime
A regime is a persistent market state — trending or ranging, calm or volatile. Strategies that work in one regime often fail in another, which is why a single backtest period proves little.
Sharpe ratio
The Sharpe ratio expresses return per unit of volatility. It is widely quoted and easy to inflate — a backtest with unusual trades or smoothed returns can produce a high Sharpe that does not survive.
Universe
The universe is the set of instruments a strategy is allowed to trade. It decides the answer before the strategy even runs — and it is where survivorship bias most often hides.
Volatility
Volatility measures how much a series moves around its average. It is not risk itself, but it is the input most risk measures are built from — and the easiest to understate.
Verification20
Alpha decay
Alpha decay is the fading of a signal's predictive power over time — after it is discovered, after it gets crowded, or simply because the edge was never stable.
Backtest
A backtest runs a trading rule over historical data to see what it would have done. It is the standard way to evaluate a strategy — and the standard way to fool yourself, if costs, look-ahead bias and survivorship are left out.
Backtest overfitting
Backtest overfitting is the specific failure of running many variants and reporting the best one. With enough attempts, a good result appears by chance alone — and looks exactly like a real one.
Backtest window
The backtest window is the period a strategy is tested over. It is the single most consequential choice in a backtest — and the one most often left unstated.
Benchmark mismatch
A benchmark mismatch is when a strategy is measured against the wrong reference — usually one that flatters it. It is easy to do, and hard to notice once it is done.
Capital capacity
Capital capacity is how much money a strategy can absorb before its own trades move the price against it. A profitable backtest says nothing about capacity — and most backtests are, in effect, run at infinite liquidity.
Confidence interval
A confidence interval gives the range a result plausibly falls in, given the noise in the data. A backtest return quoted without one is a point estimate pretending to be a fact.
Data snooping
Data snooping is the subtle form of overfitting that happens when the same dataset is used both to discover a pattern and to test it — so the test confirms only what was already fitted.
Information coefficient
The information coefficient (IC) measures how well a signal predicts forward returns — usually the correlation between a factor and the next period return. It is small in absolute terms, and that is normal.
Look-ahead bias
Look-ahead bias is when a backtest uses information that would not have been available at the time of the trade — the single most common reason an impressive backtest does not survive contact with the market.
Monte Carlo simulation
Monte Carlo simulation re-runs a strategy over many randomised paths to see the range of outcomes, not just the one that happened. It turns a single backtest number into a distribution.
Out-of-sample test
An out-of-sample test checks a rule on data it was not built on. It is the closest a backtest gets to honesty — and the step most often skipped when a result is reported.
Overfitting
Overfitting is when a rule is tuned so closely to past data that it captures noise instead of signal — and therefore fails on data it has not seen.
Paper trading
Paper trading runs a strategy against live or delayed market data without real money. It catches a class of errors a backtest cannot — but it does not test costs or psychology, and it is easy to over-read.
Point-in-time data
Point-in-time data reflects what was actually known at each moment — as first reported, with the constituents of that day. It is the opposite of a dataset quietly restated with what we know now.
Slippage
Slippage is the gap between the price a trade was expected to get and the price it actually got. It is an assumption in every backtest — and hiding it is the fastest way to make a bad rule look good.
Survivorship bias
Survivorship bias is the distortion that appears when a dataset only contains the things that survived — delisted stocks vanish from history, so a backtest over them looks better than the market actually was.
Transaction cost
Transaction cost is the total cost of making a trade — commission, spread and slippage together. Ignoring it is the quietest way to turn a losing rule into a winning backtest.
Turnover
Turnover is how much of a portfolio is replaced per period. It is the direct driver of costs — and therefore the number that decides whether a backtest survives contact with reality.
Walk-forward analysis
Walk-forward analysis re-fits a strategy on a rolling window and tests it on the next unseen period, repeating forward in time. It is closer to how a strategy is actually used than a single in-sample / out-of-sample split.