Resources

Glossary

Plain-English definitions for words that show up in the courses — grouped by career so you can skim the track you are on.

Data Science

Words you meet while cleaning tables, exploring A-level data, and writing a clear handover for a manager.

Dataset: A table of observations you analyse — rows are cases, columns are fields (for example subject entries by year and sex).
pandas: The main Python library for working with tables — load CSVs, filter rows, group, and summarise.
SQL: A language for asking questions of data stored in databases (SELECT, WHERE, GROUP BY).
Mean / median: Mean is the average; median is the middle value when sorted. Median is less pulled around by extreme outliers.
Chart: A visual summary (bar, line, box plot) that makes a pattern easier to brief than a raw table.
Handover note: A short written brief for the next person — what you found, how you got there, and what to watch out for.

Market and portfolio language used when you measure risk, test a strategy, and write an investment research note.

Return: How much a price moved over a period, usually as a percentage (daily return, annualised return).
Volatility: How much returns jump around — higher volatility means a wider range of outcomes, not “more return”.
Correlation: How two assets tend to move together (from −1 opposite, to +1 lockstep). Diversification needs less-than-perfect correlation.
Moving average: A smoothed average of recent prices (e.g. 20-day) used to spot trend without reacting to every tick.
Backtest: Run a trading rule on historical prices to see how it would have performed — useful, but not a guarantee of the future.
Buy and hold: Own the asset for the whole period with no timing trades — the usual baseline a strategy must beat after costs.
Monte Carlo: Simulate many possible future paths using random draws from a model, then summarise the range of outcomes.
Efficient frontier: The set of portfolios that offer the best expected return for each level of risk (under a given model).

Pricing and risk words used when you model claims, set premiums, and brief a risk committee.

Frequency: How often claims happen (for example claims per policy-year).
Severity: How large a claim is when it occurs (claim amount distribution).
Pure premium: Expected loss cost ≈ frequency × severity — before expenses, profit, or loading.
Loading: Extra amount added to pure premium for expenses, profit, or uncertainty.
Relativity: How much riskier (or safer) one segment is versus a base — e.g. an age-band factor of 1.3 means 30% higher expected cost.
IBNR: Incurred but not reported — claims that have happened but are not yet on the books; estimated with techniques such as chain-ladder.
Reserve: Money set aside for unpaid claims (reported and IBNR).
Reinsurance: Insurance for insurers — transfer part of large losses to another carrier for a premium.
Stress test: Replay the book under a worse scenario (flood, longevity) to see whether capital and decisions still hold.

Core modelling language for features, training, and honest evaluation before you specialise into computer vision.

Feature: An input measurement the model uses (hours studied, pixel values, a column in a table).
Label: The answer you want the model to predict (passed/failed, habitat class, clothing category).
Training set: Examples the model learns from — weights are updated on these.
Validation set: Held-aside examples used while training to check progress and tune choices (epochs, hyperparameters) — not the final exam.
Test set: Examples never used for training or tuning — evaluate once at the end for an honest score.
Epoch: One full pass through the training set.
Accuracy: Percentage of examples labelled correctly on a given set.
Overfitting: The model memorises training examples but fails on new ones — train score high, validation flat or falling.
Classifier: A model that assigns each example to a category (spam/not spam, habitat class).
Neural network: Layers of simple units (neurons) that learn weights from data; deeper stacks can represent more complex patterns.

Map-reading language for analysts who interpret model outputs and write briefs rangers can act on — without building the model.

Land-cover class: A habitat or surface label on the map (forest, grassland, water, urban, and so on).
Tile / patch: A small square cut from a larger image that the model scores as one class.
Confidence: How sure the model is about a prediction — high confidence is not the same as being right; always check the map.
Confusion matrix: Table of true labels (rows) versus predicted labels (columns). Off-diagonal cells are the mistakes that matter in the field.
Limitation: An honest bound on what the map can support (season, cloud, rare classes) — required in briefs even when accuracy looks high.
Park brief: A short note for rangers or ecologists — what the map shows, where it is weak, and what action (if any) follows.

Developer vocabulary for building and training convolutional models on satellite tiles and reporting results.

Convolution / filter: Slide a small grid of numbers over an image, multiply and sum — highlights edges or textures.
Pooling: Shrink a feature map by keeping the strongest signal in each small block.
Batch size: How many images the model looks at before updating its weights once.
Augmentation: Random flips, rotations or brightness changes during training so the model sees variation without new photos.
Backbone: A pre-trained network (e.g. MobileNetV2) that already knows edges and textures; you add a small head for your land classes.
Transfer learning: Reuse a model trained on millions of general photos, then retrain only the top for your habitats.
Softmax: Final layer that turns raw scores into class probabilities that add up to 100%.
Stratify: When splitting data, keep the same mix of classes in each part (e.g. same proportion of Forest).
Confusion matrix: Table of correct labels (rows) vs predicted labels (columns). Off-diagonal cells are mistakes.