Learn

The forecasting handbook.

Guides on demand, accuracy, drivers, models, staffing, cash, and load. Plus Predict.ai vs SageMaker and BigQuery.

Start here

Fundamentals · Start here

What is demand forecasting?

Demand forecasting estimates how much of a product or service people will want in a coming period, so you can buy, staff, and price before the rush — not after it has already walked in.

Read

Fundamentals · Start here

Time series forecasting, explained

Time series forecasting predicts the next values of something that was recorded in order — sales by day, load by hour, tickets by week — by learning from its own past, and from anything that regularly moves with it.

Read

Accuracy & evaluation · Start here

WAPE, explained

WAPE (weighted absolute percentage error) is total absolute error divided by total actuals. It behaves when some periods are zero, and it does not let tiny SKUs dominate the score the way MAPE does.

Read

In production · Start here

Why forecasts go stale in production

A forecast starts dying the moment the world moves and the model does not. New promotions, a competitor, a warehouse fire, a quiet change in how you record demand — live systems need a refresh cadence, drift checks, and a way to re-score the champion.

Read

In production · Start here

Predict.ai vs SageMaker, BigQuery ML, and Databricks

SageMaker and Databricks are workshops. BigQuery is a warehouse that can run a model. Predict.ai is a prediction OS: you name a metric, it finds drivers, races models, and keeps a live forecast with bands, scenarios, and alerts. Pick SageMaker if you need vision or a custom GPU job. That is a different product.

Read

Fundamentals

What a forecast is, and the choices before you pick a model.

Demand, time series, how far ahead to look, and how much history you need.

All in fundamentals →

Fundamentals

What is demand forecasting?

Demand forecasting estimates how much of a product or service people will want in a coming period, so you can buy, staff, and price before the rush — not after it has already walked in.

Read

Fundamentals

Time series forecasting, explained

Read

Fundamentals

Predictive analytics vs forecasting

Forecasting asks what a number will be next. Predictive analytics often asks who or what will do something. They share models and data; they do not share the clock, the score, or the decision they feed.

Read

Fundamentals

Choosing a forecast horizon

The horizon is how far ahead the forecast looks; the grain is how finely it ticks. Pick both from the decision they serve — lead time, planning cycle, and how fast the world can change — not from what the model can print.

Read

Fundamentals

How much history do you need to forecast?

Enough to see the pattern you are betting on. A weekly seasonal business wants more than one year; an hourly series can learn a week of shape faster. Quality of the target beats another dusty decade of the wrong grain.

Read

Fundamentals

Hierarchical forecasting

You rarely forecast one series. You forecast stores that must sum to a region, and SKUs that must sum to a brand. Hierarchical forecasting predicts at more than one grain and reconciles so finance and ops are not holding two official numbers.

Read

Fundamentals

Nowcasting

Nowcasting is a forecast with a horizon of this afternoon. The official close has not posted; you still have to staff the evening or dispatch the next hour. You use faster series — traffic, partial files, a weather nowcast — to estimate the present.

Read

Fundamentals

Holidays and calendar effects

A calendar is the cheapest accurate feature in the building. Christmas, payday, school terms, a 4-day week, Ramadan, Chinese New Year — if the model has to rediscover those from residuals, you are wasting folds on a date table.

Read

Accuracy & evaluation

How to score a forecast.

WAPE, MAPE, backtests, and intervals — and whether the number is good enough to plan on.

All in accuracy & evaluation →

Accuracy & evaluation

WAPE, explained

WAPE (weighted absolute percentage error) is total absolute error divided by total actuals. It behaves when some periods are zero, and it does not let tiny SKUs dominate the score the way MAPE does.

Read

Accuracy & evaluation

Choosing a forecast accuracy metric

Pick the metric that matches the cost of being wrong. WAPE for volume, MAE when units are the pain, RMSE when big misses hurt more than small ones, bias when you always run heavy or light.

Read

Accuracy & evaluation

How to backtest a forecasting model

Backtesting asks the model to forecast a past window it was not trained on, then scores the miss. Walk the window forward so you see many 'futures,' not one lucky test month.

Read

Accuracy & evaluation

What is good forecast accuracy?

Good is beating a naive baseline on the same grain, horizon, and folds — by enough to change a decision. Industry round numbers are gossip until they match your SKU mix and your clock.

Read

Accuracy & evaluation

Prediction intervals

A prediction interval is a range the future value is expected to fall in, not a promise. p50 is the middle path; p10–p90 is a band for planning stock, staff, and cash without pretending the future is a single line.

Read

Accuracy & evaluation

Forecast value added

Forecast value added (FVA) asks whether each layer — the naive, the model, the planner’s override — actually reduced error. If the override makes WAPE worse, the meeting is a cost center, not a control tower.

Read

Data & drivers

What else to feed the model.

Promos, weather, prices, leading indicators, features, and the cold-start problem when history is thin.

All in data & drivers →

Data & drivers

Forecasting with covariates

A covariate is an extra series the model is allowed to see — a promo flag, a temperature, a price. Used well, it explains swings the target's past cannot. Used as a dump of every column you have, it just overfits last quarter.

Read

Data & drivers

Finding leading indicators

A leading indicator moves before the number you care about. Search queries before orders, bookings before departures, parts receipts before output. The lag is the product; correlation on the same day is just a mirror.

Read

Data & drivers

Feature engineering for time series

Most of the work is turning a clock and a few raw series into lags, rolling stats, and calendar flags the model can chew. Do that on a schedule, with no future leaking into the past, or the backtest is theater.

Read

Data & drivers

Forecasting new products and locations

A cold start is a forecast with almost no history of its own. You borrow shape from similar items or places, use leading signals, keep the horizon short, and replace the borrowed prior as soon as real sales show up.

Read

Data & drivers

Intermittent demand

Spare parts, rare claims, cyber events, slow B2B orders — the series is mostly zeros, then a jump. Average error metrics lie. You often care about hit rate and size when a hit happens, not a daily mean of 0.4 units.

Read

Data & drivers

Driver discovery

A driver is a series that improves the forecast when you are allowed to see it — lagged, tested, and still available at 6am. Correlation on a heatmap is not discovery. Discovery is a backtest with the candidate in and out.

Read

Data & drivers

Time series data quality

Most forecast failures are not the model. They are a DST shift, a file of zeros that looked like a miracle, sales used as demand after stockouts, or a timestamp that is actually an invoice date. Fix the clock before you race architectures.

Read

Data & drivers

Price, promo, and lift

Lift is what happens to the path when you change a price or a promo you actually control. It is a scenario on a model that has seen similar moves — not a 2.4 elasticity copied from a textbook onto a fare class, a kWh tariff, or a SKU.

Read

Models & methods

Baselines through foundation models.

What each method is for, what zero-shot changes, and why production keeps more than one model in the race.

All in models & methods →

Models & methods

Forecasting models compared

Classical models are interpretable and cheap. Machine learning eats more drivers. Foundation models arrive pretrained and can go zero-shot. None of them win every series — which is why serious stacks keep a tournament, not a favorite.

Read

Models & methods

Zero-shot forecasting

Zero-shot means you get a forecast without training on your series first. A foundation model uses what it learned from many other series, then reads yours. Fine-tuning is the later step, once you have enough of your own data to make it worth the spend.

Read

Models & methods

Why one model never wins

A model that dominated last spring can lose in November. Series change, drivers change, and luck exists. Keep a challenger on the same folds, promote on evidence, and stay loyal to the score — not to the architecture that won a bake-off in March.

Read

In production

After the forecast ships.

Stale models, drift, serving, alerts, what-ifs, and handing the loop to agents.

All in in production →

In production

Why forecasts go stale in production

Read

In production

Anomaly detection for business metrics

A fixed threshold treats every Tuesday like every Saturday. A forecast band asks whether today is weird given what the series usually does at this clock. That is the difference between an on-call fire drill and a useful alert.

Read

In production

What-if analysis for forecasts

A what-if is a second forecast under a change you control — price up 8%, a promo pulled, a plant offline. You compare paths, not opinions. The point is not a perfect simulation; it is catching the decision that only looks cheap this week.

Read

In production

Forecasting with AI agents

An agent can create a forecast, ask why it moved, and push the number into the tools you already run — if the platform exposes those jobs as tools, not as a chat box glued to a screenshot. The forecast still has to be a real model with a score.

Read

In production

Serving forecasts through an API

A forecast that lives in a slide is a report. A forecast that answers a request — grain, horizon, quantiles — can staff a roster, bid a market, or fill a warehouse system. The hard part is the contract: version, pin, and what happens when the model changes.

Read

In production

S&OP and the forecast

Sales and operations planning wants a number the company can staff, buy, and book against. It fails when each function brings a private forecast. The work is not a prettier chart. It is one grain, one horizon per decision, and a cadence that matches the meeting.

Read

In production

Predict.ai vs SageMaker, BigQuery ML, and Databricks

Read

Decisions

Staff, cash, load, beds, claims.

The clocks other teams live on — rosters, megawatts, cash, occupancy, and the S&OP meeting.

All in decisions →

Decisions

Workforce and staffing forecasts

Nurses, agents, cooks, field techs — the cost is people on a clock. A staffing forecast is arrivals or contacts by interval, with a band wide enough for a call-out list, not a single headcount that makes Tuesday look like Saturday.

Read

Decisions

Energy load forecasting

Load, generation, and price live on a tight clock. Weather is a covariate you only get as a forecast. A skinny band at hour 36 is swagger. Operators plan reserves on a high quantile, not on a pretty MAE.

Read

Decisions

Cash flow forecasting

Treasury cares when money lands, not when revenue was recognized. A cash forecast is collections, payroll, tax, and drawdowns on a calendar — messy, lumpy, and allergic to MAPE if a day can be negative or zero.

Read

Decisions

Revenue forecasting

Revenue is not one series. Bookings, billings, recognized revenue, and pipeline coverage all have different clocks. SaaS, usage, and store sales only look alike in a board deck. Pick the definition the decision uses, then forecast that.

Read

Decisions

Capacity planning

A plant, a ward, a classroom, a network link — capacity is a ceiling. The forecast is whether you will hit it, and when. Over-build is capital. Under-build is a queue, a blackout, or a denied admission.

Read

Decisions

Claims and incident volume

Fraud queues and claim shops live on volume: how many will land next week, by type and region. That is a forecast. Who is risky is a score. Mixing AUC into a staffing plan is how you get a confident dashboard and an overtime surprise.

Read

Decisions

Occupancy and travel demand

A hotel night and a flight seat perish at midnight. The forecast is pickup along a booking curve, not only yesterday’s occupancy. Price is a lever you set; remaining capacity is the constraint. Those two belong in the same conversation.

Read

Decisions

Patient flow forecasting

Clinics, EDs, and wards are queues with clinical consequences. Forecast arrivals by hour and service, not a daily average that hides the 11am surge. Staffing and beds are downstream of that path.

Read

Decisions

Lead time forecasting

Suppliers, customs, labs, and work orders all have a duration. That duration moves. Forecasting lead time is how you stop treating '14 days' as a constant in an MRP that has not been true since March.

Read

Decisions

FP&A forecasting

Finance needs a slower, cleaner number than a warehouse. The mistake is either ignoring the operational forecast or forcing it into a 12-month line that cannot be true. FP&A should consume a reconciled path, plus scenarios, plus an interval.

Read

Glossary

WTF is WAPE? A lag? A quantile? Short answer, then a link if you want more.

Browse all 36 terms →

Time series Seasonality Trend Stationarity Forecast horizon Granularity Covariate Exogenous variable Lag feature Rolling window Backtesting Walk-forward validation WAPE MAPE sMAPE MAE RMSE MASE

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.

Start with your data Talk to sales