Glossary

The words this course runs on, one line each, with the week where each one becomes a number you compute.

You will not learn these from this page. You will learn each one the first time it costs you a wrong number. This page is for when you forget which week that was.

The Wk column says where the term is introduced properly.

Language models

Term	In one line	Wk
LLM (large language model)	A model trained to predict the next piece of text, given all the text before it. Everything else it does is built on that one operation.	1
Neuron	Multiply each input by a weight, add them up with a bias, squash the result. One neuron with a sigmoid squash is a logistic regression.	1
Neural network	Neurons in layers, each layer feeding the next. The hidden layers invent their own features; nobody tells them which.	1
Parameters	All the weights and biases of a model — the numbers it learns. “117B parameters” means 117 billion of them.	1
Training (gradient descent)	Measure how wrong the model is (the loss), nudge every weight slightly in the direction that reduces it, repeat — millions of times.	1, 3
Pre-training	Training a network to guess the next token on a huge amount of text. Produces a base model, which continues documents rather than answering questions.	1
Post-training	Turning a base model into an assistant: supervised fine-tuning (SFT) on example conversations, then reinforcement learning (RL, RLHF) on what gets rated well.	1
Transformer	The architecture of every current LLM: a block of attention plus neurons, repeated dozens of times.	1
Mixture-of-experts (MoE)	A model split into many specialist sub-networks, of which only a few run for each token — the capacity of a big model at the running cost of a small one.	1
Quantization	Storing each weight in fewer bits (e.g. 4 instead of 16). Smaller and cheaper to run, slightly less precise.	1
Open-weight	The weights can be downloaded and run on your own machines. Not the same as open-source: the training data and code usually stay closed.	1
Token	The unit a model reads and writes: a chunk of characters, often part of a word. You are billed, limited and measured in tokens, not words.	1
Tokenizer	The component that cuts text into tokens. It is trained on a corpus, so text unlike that corpus (other languages, numbers, new names) cuts into more pieces.	1
Context window	Everything the model can see in one call: instructions, history, documents, and its own answer so far. Nothing outside it exists for the model.	1
Attention	How the model reads its context: to process each token it weighs every other token in the window, and decides how much each one matters. Every-token-with-every-token is why long contexts cost more than they look.	1
Weights	The numbers learned in training. They are frozen when you call the model; nothing you type changes them.	1
Prompt	The text you send. In production it is code: versioned, tested, and stored next to what it produced.	1–4
Temperature	How strictly the model follows its own probabilities when choosing the next token. T = 0 for extraction and anything auditable; higher for ideas.	2
Hallucination	A fluent output that is not supported by anything in the context — the model producing the shape of an answer where the facts are missing.	2
Embedding	A list of numbers representing a piece of text, arranged so that similar meanings are close together. Turns “similar” into a distance.	4
RAG (retrieval-augmented generation)	Find the relevant passages in your documents first, put them in the context, then ask the model to answer from them — and cite them.	4
Structured output	Forcing the answer into a fixed shape (e.g. JSON with named fields). Guarantees the type, not the truth.	4
Fine-tuning	Further training of a model on your examples. Changes style and format well; is a poor way to add facts you will need to update or cite.	1, 4
Tool call	The model emits a structured request (“call get_price('BTC')”); your code runs it and returns the result into the context.	5
Agent	A model calling tools in a loop until it decides it is done. Every agent failure is a failure of that loop.	5
Prompt injection	Instructions hidden inside a document or web page that the model reads and then obeys.	5, 9
Eval (evaluation)	A fixed set of questions with known answers, run on every change, so “it seems better” becomes a number.	5
LLM-as-judge	Using one model to grade another’s answers. Useful and cheap; biased, so it is itself checked against human grades.	5
Software 1.0 / 2.0 / 3.0	The same decision written as a rule, as a model fitted to data, or as an instruction in plain language. Knowing which one a problem needs is most of the job.	1, 3

Data and analysis

Term	In one line	Wk
Notebook	A document mixing code, its output and prose, run cell by cell. Ours run in Google Colab.	1
DataFrame	A table in Python (pandas), with named columns and an index — usually dates, for us.	2
Return (simple)	Percentage change from one price to the next: P_t / P_{t-1} − 1.	2
Log return	ln(P_t / P_{t-1}). Unlike simple returns, log returns add up over time.	2
Cumulative return	Total growth over a period. With simple returns it is a product of (1 + r), not a sum.	2
Volatility (annualised)	Standard deviation of daily returns, scaled to a year by √(trading days): 252 for stock markets, 365 for crypto.	2
Drawdown	The fall from the highest value reached so far. Maximum drawdown is the worst such fall in the period.	2
Join	Putting two tables side by side on a shared key (for us, the date). The choice of which rows survive is a modelling decision.	2
Forward-fill (ffill)	Filling a missing value with the last known one. It invents data — sometimes that is the right choice, always it must be said.	2
Units	What a number measures: percent or decimal, dollars or thousands. Not in the data; always your responsibility.	2
Snapshot	A dated copy of the data behind a result, so the number can be reproduced when the source changes.	2
Reproducibility	Someone else, on another machine, gets your number. If it needs you in the room, it is not a result yet.	2

Machine learning and evaluation

Term	In one line	Wk
Feature	What you know at time t, used as input.	3
Target	What happens after t — the thing you are trying to predict.	3
Horizon	How far after t the target is measured.	3
Base rate	How often the outcome happens anyway. Predicting “always the most common outcome” scores the base rate with no model at all.	3
Baseline	The simplest reasonable alternative your model must beat — the base rate, a rule, last week’s value.	3
Train / test split	Fit on one part of the data, measure on another the model never saw.	3
Leakage	Information from the future (or from the test set) getting into what the model learns from. Produces results that look excellent and are worthless.	3
Walk-forward	Repeatedly fit on the past and test on the next period, moving forward in time. The honest split for anything with a date on it.	3, 7
Accuracy	Share of predictions that were right. Meaningless without the base rate.	3
Precision / recall	Of what you flagged, how much was right / of what was there, how much you flagged. You trade one for the other with the threshold.	3
AUC	How well a model ranks positives above negatives, across all thresholds. 0.5 is a coin.	3
Threshold	The probability above which you act. Set it from the costs of the two kinds of error, not at 0.5 by habit.	3
Calibration	Whether “70% likely” happens about 70% of the time. A model can rank well and be badly calibrated.	3
SHAP	A method that splits a single prediction into contributions from each feature. Explains the model, not the world.	9

Markets and backtesting

Term	In one line	Wk
Backtest	Running a trading rule on historical data as if you had traded it. The easiest result in finance to fool yourself with.	7
Transaction costs	Fees, spread and price impact paid on every trade. Many strategies that work on paper do not survive them.	7
Survivorship bias	Testing only on assets that still exist today, which quietly removes the ones that failed.	7
Multiple testing	Try enough strategies and one will look good by chance. The more you tried, the less the best one means.	7
Regime change	The market’s behaviour shifts (rates, volatility, regulation), and a pattern that held stops holding.	7
Sharpe ratio	Average excess return divided by its volatility — return per unit of risk.	7
Model risk	The risk of loss from a model that is wrong, or used where it was not meant to be. Banks are required to manage it.	7

Crypto and DeFi

Term	In one line	Wk
Blockchain	A shared ledger that many independent computers keep in agreement, where past entries are very hard to change.	8
On-chain data	Everything recorded on the ledger — transfers, balances, contract calls — public, and readable by anyone.	2, 8
Wallet / address	An address holds assets; a wallet holds the private key that can move them. Whoever has the key has the assets.	8
Smart contract	A program deployed on a blockchain that runs exactly as written when called. Including its bugs.	8
Gas	The fee paid to the network to execute a transaction or a contract call.	8
Stablecoin	A token designed to hold a fixed value, usually one US dollar (USDC, USDT). Only as stable as what backs it.	8
DeFi (decentralised finance)	Financial services — trading, lending, borrowing — run by smart contracts instead of intermediaries.	8
DEX (decentralised exchange)	An exchange run by a smart contract. Most use a liquidity pool instead of an order book.	8
Liquidity pool	Two tokens deposited in a contract that anyone can trade against.	8
AMM (automated market maker)	The rule that prices a pool. The classic one: x · y = k — the product of the two balances stays constant, so every trade moves the price.	1, 8
Slippage / price impact	The difference between the quoted price and what you actually pay, because your own trade moves the price.	1, 8
LP (liquidity provider)	Someone who deposits tokens into a pool and earns a share of trading fees.	8
Impermanent loss	What an LP gives up, compared with simply holding the two tokens, when their relative price moves. “Impermanent” only if the price comes back.	8
TVL (total value locked)	The value of assets deposited in a protocol. A size measure, not a profit measure — and easy to inflate.	8
Lending protocol	A contract that lends deposited assets to borrowers who post more collateral than they borrow.	8
Collateral ratio / liquidation	How much collateral backs a loan. Fall below the threshold and the protocol sells your collateral automatically.	8
Staking	Locking tokens to help secure a network (or a protocol) in exchange for a yield.	8
Oracle	A service that brings off-chain data (e.g. a price) onto the chain. A wrong or manipulated oracle price is a classic DeFi failure.	8

Regulation and responsible use

Term	In one line	Wk
EU AI Act	EU regulation that classifies AI uses by risk — prohibited, high-risk, limited, minimal — with obligations rising accordingly. Credit scoring is high-risk.	4
MiCA	EU Markets in Crypto-Assets regulation: licensing and rules for crypto-asset service providers and stablecoin issuers.	8
ESMA	The EU securities markets authority; issues guidance under MiCA and on AI in investment services.	8
GDPR	EU data protection law. For us: personal data does not go into an AI tool without a legal basis — in this course, not at all.	2, 9
Human oversight	A named person who can understand, question and override the system’s output — and actually does.	9
Red-teaming	Attacking your own system on purpose to find how it fails before someone else does.	9