GLOSSARY · QUANT & AI · FOR NON-SPECIALISTS25 TERMS · 30-SECOND READS · DEEP-LINKABLE
CHRISTIAN.T.MACIONUTC+825 TERMSDEEP-LINKABLE
A short dictionary of the terms on this site.
Twenty-five terms, each short enough to read in thirty seconds, each deep-linkable from /methodology.
Twenty-five canonical definitions. Each one is short enough to read in thirty seconds and stable enough to deep-link to from the methodology, proof, or solutions pages.
Quant terms · 17
Statistical-evaluation vocabulary. Where possible, terms link to the methodology page where they are applied in practice.
- →Alpha (α)
- The portion of an investment's return that is not explained by exposure to broad market risk. The signal beyond the benchmark.
- →Block bootstrap
- A resampling method that preserves short-term autocorrelation in time-series by sampling contiguous blocks rather than individual data points. Used to build honest confidence intervals.
- →Bonferroni to Holm correction
- A multiple-testing correction applied when many hypothesis tests are run at once. Prevents the probability of any false positive from inflating as the number of tests grows.
- →Cointegration
- A statistical property of two or more time-series that move together in the long run even though each one individually wanders. The basis for pairs and stat-arb strategies.
- →CSCV / PBO
- Combinatorially Symmetric Cross-Validation, the standard method for estimating Probability of Backtest Overfitting. Tells you how many of your backtest winners would have been selected by chance alone.
- →Deflated Sharpe Ratio (DSR)
- A correction to the Sharpe ratio that adjusts for the number of trials, the distribution of returns, and the skew/kurtosis of the strategy. Tells you whether a high Sharpe is real or a multiple-testing artifact.
- →Drawdown (DD)
- The peak-to-trough decline of an equity curve over a specified window. The most-cited measure of risk in a systematic book.
- →Embargo
- A gap between the train set and the test set in walk-forward evaluation. Prevents leakage of recent information into the model used for older data.
- →G1 to G31 (evaluation gates)
- A 31-gate statistical evaluation stack applied to every quantitative project on this site. Covers leakage, multiple-testing, walk-forward, DSR, PBO, transaction-cost modelling, and OOS paper-trade.
- →MinBTL (Minimum Backtest Length)
- The minimum number of trades a backtest must contain before its Sharpe ratio is statistically distinguishable from zero at a given confidence level.
- →Out-of-sample (OOS)
- Data the model has never seen during training or parameter selection. The closest a backtest gets to a real test of generalisation.
- →PBO (Probability of Backtest Overfitting)
- The probability that the best backtest winner, selected by in-sample performance, underperforms the median out-of-sample. Estimated by CSCV.
- →Regime
- A persistent state of the market (high-vol, low-vol, trending, mean-reverting, risk-on, risk-off) that affects which strategies work and which do not.
- →Sharpe ratio
- The average excess return of a strategy divided by its standard deviation. The canonical risk-adjusted return measure.
- →Slippage
- The difference between the expected fill price of a trade and the price at which it actually executes. A major component of transaction cost in liquid markets.
- →Survivorship bias
- A dataset error where only assets that "survived" to the present are included, biasing the historical sample toward winners. Common in equity-index backtests.
- →Walk-forward evaluation
- A rolling evaluation where the model is retrained on a moving window and tested on the immediately following window. The most honest single-shot backtest.
AI terms · 8
Multi-agent, retrieval, and evaluation vocabulary used across the AI lane of this site.
- →Agent charter
- A short document that defines an AI agent's job, inputs, outputs, and failure modes before it is built. Every agent on this site ships with one.
- →Eval harness
- A test rig that runs a model or agent through a fixed set of inputs, scores the outputs against a rubric, and persists the scores for trend analysis. The AI equivalent of a quant backtest.
- →Frozen spec
- A pinned version of a model, prompt, and tool set used inside an eval. The spec is immutable for the duration of the eval so scores are reproducible.
- →JSON Schema
- A declarative specification for the shape of a JSON document. Used as a contract between agents and as a validator inside eval harnesses.
- →LLM-as-judge
- Using a language model to grade the outputs of another model on dimensions that are hard to express as a deterministic check (tone, completeness, faithfulness).
- →MCP (Model Context Protocol)
- A protocol for connecting language models to tools, data sources, and other agents over a typed JSON-RPC interface. The eval-mcp-server on this site conforms to MCP 2025-06-18.
- →Multi-agent system
- A system composed of multiple specialised agents that coordinate to complete tasks a single agent could not. The orchestrator-worker pattern is the dominant topology.
- →RAG (Retrieval-Augmented Generation)
- A pattern where a language model is given retrieved context (chunks from a vector store) before generating its answer. Reduces hallucination on factual queries.
See these terms in action.
Each definition above links to a method or artifact on this site where it is applied. The methodology page maps every gate (G1 to G31) to the term it implements.