Appendix C · Glossary

15.1 Cross-disciplinary terminology

This glossary maps the operational vocabulary between two disciplines. For the quant, the first table explains the LangChain and LangGraph components that govern financial agents. For the LLM engineer, the second defines the quantitative concepts that shape signals, portfolios, risk, and execution. Cross-references indicate the module where each term receives practical treatment.

15.1.1 40+ terms in two tables

Table C.1 — LLM/LangChain engineering

Term	Concise definition	Module where it is covered
agent	System that uses an LLM to iteratively decide actions, invoke tools, and update its state until a task is complete.	Module 3
checkpoint/checkpointer	State persisted at each graph step; enables recovery, resumption, replay, and auditable forking of executions.	Modules 3, 11, and 12
content blocks	Typed, normalized units of a message —text, image, audio, or tool reference— independent of the provider.	Module 2
context_schema	Runtime context schema that injects typed data into the agent without mixing it with conversational state.	Module 3
create_agent	Standard LangChain v1 factory that composes model, tools, prompt, and middleware into a compiled LangGraph graph.	Module 3
embedding	Numeric vector representing text or data; geometric proximity approximates semantic similarity for search and clustering.	Module 5
grounding	Anchoring of claims and figures to verifiable documents or tools; reduces hallucinations but does not guarantee accuracy.	Modules 6 and 11
guardrail	Validation control applied before, during, or after the model that enforces format, safety, PII, or grounding policies.	Module 11
HITL (human-in-the-loop)	Human oversight integrated into the flow to approve, edit, or reject materially relevant actions before execution.	Modules 3, 10, and 12
interrupt	Controlled LangGraph pause that saves state and waits for human or external input via `resume`.	Modules 3 and 12
LangGraph	Orchestration framework for stateful flows with cycles, persistence, subgraphs, recovery, and human control.	Module 3
LangSmith	Platform for traces, datasets, evaluations, experiments, dashboards, and alerts to observe agents in development and production.	Module 11
LCEL/Runnable	LangChain's declarative composition model: units with `invoke`, `batch`, and `stream` connected via the \| operator.	Module 2
MCP (Model Context Protocol)	Open client-server protocol for exposing external tools, resources, and prompts to LLM applications in a uniform way.	Modules 3 and 4
middleware	Layer that intercepts model or tool calls to summarize, retry, filter PII, cap cost, or require approval.	Module 3
RAG	Retrieval-Augmented Generation: retrieves relevant context before generating, separating document memory from model reasoning.	Module 5
reducer	Function that combines concurrent or successive updates to a state channel, for example appending messages without overwriting them.	Module 3
state/schema	The graph's data contract: defines fields, types, and update rules shared by nodes, tools, and checkpoints.	Module 3
store	Persistent memory across threads or conversations; retains theses, preferences, and knowledge retrievable by keys.	Modules 3 and 12
structured output	Output validated against a schema —typically JSON or Pydantic—; ensures parseable shape, not factual truth.	Modules 2 and 6
subgraph	Compiled graph embedded as a node in another graph; encapsulates a team or process while keeping state and streaming observable.	Modules 3 and 12
token	Atomic unit of model processing and billing; prompts, retrieved context, and responses consume token budget.	Modules 2 and 11
tool calling	Mechanism by which the model emits a structured call to a function; execution remains deterministic and external.	Modules 2 and 6
vector store	Specialized store that indexes embeddings and metadata to retrieve documents by similarity, filters, and hybrid ranking.	Modules 4 and 5

Table C.2 — Quantitative finance

Term	Concise definition	Module where it is covered
alpha/alpha decay	Alpha is excess return not explained by systematic risk; its decay is the erosion from diffusion, crowding, costs, or loss of relevance.	Modules 4 and 9
backtesting	Historical simulation of a strategy with the data, costs, and constraints available ex ante; estimates performance without guaranteeing future replication.	Module 9
Black-Litterman	Bayesian portfolio model that combines market equilibrium, $\pi=\delta\Sigma w_{mkt}$ , with investor views weighted by uncertainty.	Module 8
borrow	Availability and cost of borrowed securities to maintain shorts; includes locate, fee, rebate, and recall risk.	Module 9
CVaR/Expected Shortfall	Average loss conditional on exceeding VaR; measures the tail better than the isolated quantile, and FRTB uses ES at 97.5%.	Module 7
DSR (Deflated Sharpe Ratio)	Sharpe adjusted for number of trials, sample variance, and non-normality; estimates whether the result exceeds selection bias.	Modules 8 and 9
EWMA	Exponentially weighted moving average; daily RiskMetrics uses $\lambda=0,94$ to give more weight to recent observations.	Module 7
factor (Fama-French)	Systematic source of return; the 3/5-factor models include market, size, value, profitability, and investment.	Module 8
FRTB	Basel market risk framework: ES at 97.5%, liquidity horizons, non-modellable factors, and limits on internal models.	Module 7
GARCH	Conditional volatility model; in GARCH(1,1), $\sigma_t^2=\omega+\alpha\varepsilon_{t-1}^2+\beta\sigma_{t-1}^2$ .	Module 7
HRP	Hierarchical Risk Parity: clusters assets by correlation and allocates via recursive bisection without inverting the covariance matrix.	Module 8
Kelly	Fraction of capital that maximizes expected logarithmic growth; continuous approximation $f^*=(\mu-r_f)/\sigma^2$ , with fractional Kelly for robustness.	Module 8
look-ahead bias	Bias from using information unavailable at decision time, such as revised fundamentals or later timestamps.	Modules 4 and 9
market impact (Almgren-Chriss)	Cost of moving the price when executing; the model separates temporary and permanent impact and balances cost against execution risk.	Module 9
MNPI	Material non-public information whose use in prompts, tools, or decisions can trigger insider dealing and compliance risk.	Module 10
PBO	Probability of Backtest Overfitting; via CSCV it estimates the probability that the best in-sample configuration underperforms out of sample.	Module 9
point-in-time	Data preserved as it was known at each instant, with universes, revisions, and effective dates without later rewriting.	Modules 4 and 9
risk parity/ERC	Allocation that equalizes contributions to total risk; ERC requires identical $w_i(\Sigma w)_i$ for all assets.	Module 8
Sharpe/Sortino	Sharpe divides excess return by total volatility; Sortino penalizes only volatility below a threshold or target return.	Module 7
slippage	Difference between expected and executed price due to latency, spread, impact, or liquidity; turns gross profits into net.	Module 9
survivorship bias	Bias from analyzing only securities or funds that survive, omitting bankruptcies, mergers, and delistings from the historical universe.	Modules 4 and 9
tracking error	Standard deviation of active return versus a benchmark; quantifies the consistency of the deviation, not necessarily its sign.	Module 7
VaR	Loss quantile at a given confidence and horizon —for example 99% over one day—; it does not represent the maximum possible loss.	Module 7
human veto	Operational right to stop, modify, or reject a recommendation or order before execution; it closes the accountability loop.	Modules 1, 9, and 12
cross-sectional z-score	Standardization $(x_i-\bar{x}_t)/s_t$ within the universe on the same date; makes signals comparable and reduces scale effects.	Module 9

Appendix C of 3 — LangChain for Quantitative Trading. Data and versions verified as of July 2026; market prices and figures subject to change. LangChain 1.3.14 / langchain-core 1.5.2.