WTF is WAPE?
A lag, a quantile, cold start — one-paragraph answers. Open a term if you want the full guide.
Fundamentals
Forecast horizon
The forecast horizon is how far into the future the model is asked to fill in. A 14-day horizon is fourteen daily ticks (or two weeks). It should match the decision's lead time, not a default in a library.
Granularity
Granularity is the size of each tick: hourly, daily, weekly, SKU-store versus category-region. Finer grain looks more detailed and is usually noisier. Pick the grain the planner actually uses.
Nowcasting
Nowcasting estimates the present (or the very next tick) using faster series, because the official number has not landed yet. It is a forecast with a horizon of 'this afternoon,' not next quarter.
Seasonality
Seasonality is a pattern that repeats on a fixed clock — weekly, yearly, intraday — because people, weather, and operations do. It is not a trend, and it is not a one-off spike.
Stationarity
A series is stationary when its mean and variance sit still over time. Many classical models assume that. Real business series usually do not — they trend, they season, they jump when you change the website.
Time series
A time series is a sequence of measurements recorded in order — one value per timestamp — so yesterday can inform tomorrow in a way a shuffled table of the same numbers cannot.
Trend
Trend is the slow drift in level — up, down, or flattening — after you look past the weekly wiggle. Growth, decay, saturation. It is the part of the series that would still be there if you blurred out the season.
Accuracy & evaluation
Backtesting
Backtesting scores a forecast on a past window the model was not trained on. Done properly, you walk that window forward so you see many futures, not one lucky test month.
Baseline forecast
A baseline is a simple rule you must beat: last value, last week, last year, a seasonal naive, a moving average. If a fancy model cannot beat it on a fair backtest, keep the rule and save the drama.
Forecast bias
Bias is the average signed error: systematically high or systematically low. A forecast can have a handsome WAPE and still drain cash if it is always 6% heavy.
MAE
MAE is the average of |forecast − actual| in the original units. If every missed unit costs about the same, it is the everyday score that matches the cost.
MAPE
MAPE averages |error / actual| across periods. It is popular, easy to quote, and badly behaved when actuals are near zero or when small items should not outvote large ones.
MASE
MASE divides your MAE by the MAE of a naive seasonal baseline. Below 1 means you beat 'same as last week' (or last year). It is how you compare series of different sizes without a percent.
Prediction interval
A prediction interval is a range a new observation is expected to fall in. It is not a confidence interval around a mean, and it is not a guarantee. p10–p90 is a planning band, not a promise.
Quantile forecast
A quantile forecast predicts a percentile of the future distribution — p50 as the middle, p90 as a high plan — instead of (or in addition to) a single expected value.
RMSE
RMSE is the square root of the average squared error. Big misses hurt more than small ones. Use it when a rare large error is worse than many small ones — not because it looks more 'statistical.'
sMAPE
sMAPE is a percentage error that divides by the average of |actual| and |forecast|, so zeros on one side do not explode as violently as MAPE. It is still a percent, and it still is not a volume-weighted score.
Walk-forward validation
Walk-forward validation is backtesting that respects time: train on the past, predict the next horizon, slide the origin forward, repeat. You never train on a Wednesday to predict the Tuesday before it.
WAPE
WAPE is the sum of absolute errors divided by the sum of actuals, usually shown as a percent. It weights misses by volume, so a tiny SKU cannot dominate the score the way MAPE allows.
Data & drivers
Cold start
A cold start is a forecast for a series with little or no history of its own — a new SKU, a new store, a new region. You borrow shape from analogs until the series can stand up.
Covariate
A covariate is an extra series the forecast is allowed to see besides the target — price, promo, weather, spend. It only helps if you will have it (or a forecast of it) at prediction time.
Exogenous variable
In forecasting jargon, exogenous means 'coming from outside the target series.' Price, weather, and marketing spend are typical. The word does not make the variable good. Availability at forecast time does.
Feature engineering
Feature engineering turns raw timestamps and series into the lags, rolling stats, and calendar flags a model can use. In time series, the unforgivable sin is building those features with the future mixed in.
Hierarchical forecasting
Hierarchical forecasting predicts at more than one grain — SKU, category, chain — and reconciles so the pieces add up. It is how you keep a detailed plan that still matches the number finance already believed.
Intermittent demand
Intermittent demand is a series that is zero most of the time and jumps when it is not — spare parts, slow SKUs, B2B orders. Average error metrics lie; you often care about hit rate and size when a hit happens.
Lag feature
A lag feature is a past value of a series, shifted by k steps, used as an input. Lag 7 on daily data is 'the same weekday last week.' Most time-series models are careful lag machines.
Leading indicator
A leading indicator moves before the target. Clicks before orders, bookings before departures, parts in before units out. Same-day correlation is a mirror; the lag is the product.
Rolling window
A rolling window is a moving slice of the last n points used to compute a statistic — mean, max, standard deviation. It summarizes recent level without keeping every lag as its own column.
Models & methods
Champion–challenger
Champion–challenger means one model is live (the champion) while others train on the same data and folds (challengers). If a challenger wins by enough, it takes the job. Loyalty is to the score.
Ensemble
An ensemble combines more than one model — an average, a median, a weighted stack — so no single architecture has to be right on every week. Tournaments that pick a winner are a hard ensemble; blending is a soft one.
Fine-tuning
Fine-tuning continues training a pretrained model on your series (or your workspace) so the general prior specializes. It costs more than zero-shot and should earn that cost on a backtest.
Foundation model
A forecasting foundation model is pretrained on many time series, then used zero-shot or fine-tuned on yours. The bet is that other series taught it shapes that transfer — season, spikes, slow trends — so you are not starting from random weights.
Zero-shot forecasting
Zero-shot forecasting produces a path for a series the model was not trained on, using a pretrained foundation model. You pass history at inference time. There is no local training job on that series first.
In production
Anomaly detection
Anomaly detection flags values that do not match what the series usually does at that clock. On seasonal business metrics, a forecast band beats a static threshold that treats every Tuesday like every Saturday.
Data drift
Data drift is the inputs changing shape — new POS codes, a missing promo feed, a shifted mean. Concept drift is the relationship changing. Both make last month's champion quietly worse.
What-if analysis
What-if analysis is a second forecast under a change you control — price, promo, capacity. You compare paths from the same model, not two opinions in a spreadsheet.
Try it on your data.
Connect an outcome and its history. Predict.ai finds the drivers, races the models, and keeps the forecast live — in the workspace, over the API, and through agents.