AI Forecasting Software Development: Architecture, Data and Model Evaluation
How to build AI forecasting software: data, models, accuracy metrics, backtesting, planner workflows, and how it differs from BI and decision intelligence.
The short answer
AI forecasting software predicts future values such as demand, sales, cash flow or staffing needs from historical data and known drivers, and gives planners a range of likely outcomes rather than a single number. Building it well depends less on the choice of model than on clean, joined data, honest evaluation against simple baselines on past periods, forecasts at the level decisions are made, and a workflow where planners can review, adjust and explain the numbers before they flow into ERP and planning systems.
This guide is for operations, finance and supply chain leaders and the technology teams supporting them. It explains how forecasting differs from BI, predictive analytics and decision intelligence, what data you need, how to evaluate models, the architecture, and a phased build. No forecast is always accurate; good software makes uncertainty visible and useful.
BI, predictive analytics, forecasting and decision intelligence
These terms overlap, and vendors use them loosely. The practical difference is the question each one answers.
| Category | Question it answers | Typical output | Maturity |
|---|---|---|---|
| Business intelligence (BI) | What happened, and where? | Dashboards and reports on historical data | Mature, widely used |
| Predictive analytics | What is likely to happen to this customer, order or asset? | Scores and classifications (churn risk, fraud risk, failure risk) | Mature |
| Forecasting | How much of something will there be over time? | Time series of future values with ranges (units per week, cash per month) | Mature methods; AI models adding to them |
| Decision intelligence | Given the forecasts and constraints, what should we do, and what happens if we do? | Recommended actions, scenarios, decision logs | Emerging as a product category; often assembled from the layers above |
Worth noting
Decision intelligence is a useful framing, popularised by industry analysts; Gartner published its first Magic Quadrant for decision intelligence platforms in January 2026. Most of the value still comes from the layers beneath it: reliable data, good forecasts and clear decision rules. Build those first.
What businesses forecast
| Use case | Typical level | Drivers worth including |
|---|---|---|
| Demand and replenishment | Product × location × week (or day) | Promotions, prices, holidays and festivals, stockouts, weather where relevant |
| Sales and revenue | Region or channel × month | Pipeline, seasonality, pricing changes, marketing spend |
| Cash flow | Entity × week | Invoice due dates, payment behaviour, payroll and tax dates |
| Workforce and capacity | Site or team × hour or day | Order volumes, appointment bookings, seasonal patterns |
| Production planning | Product family × week | Orders, lead times, capacity, supplier reliability |
Data comes first
Most forecasting projects spend more time on data than on models. Forecasts are only as good as the history they learn from, and sales history usually hides problems:
- Stockouts and lost sales: zero sales when the item was out of stock is not zero demand. Flag these periods or the model will learn to under-forecast.
- Promotions and price changes: record them with dates, depth and channel, or spikes will look like random noise.
- Product changes: new products, replacements and discontinued lines need mapping so history carries over sensibly.
- Calendar effects: local holidays and festivals that move each year (for example Diwali, Eid or Easter) and month-end or payday effects.
- Returns and cancellations: decide whether you forecast gross orders or net demand, and be consistent.
- Granularity and hierarchy: product, category, location and channel hierarchies. Forecasts made at different levels rarely add up on their own; reconciliation methods make them consistent, and usually improve accuracy at most levels.
Models: start simple, then earn complexity
Forecasting has a long, well-tested toolbox, and newer AI models add to it rather than replace it. A sensible progression:
| Approach | Strengths | Limitations |
|---|---|---|
| Simple baselines (last value, seasonal naive, moving average) | Fast, transparent; the benchmark every other model must beat | Ignore drivers and changing patterns |
| Statistical models (exponential smoothing, ARIMA) | Strong on stable, seasonal series; produce ranges naturally | One model per series; limited use of external drivers |
| Machine learning (gradient-boosted trees and similar) | Learn across many products at once; use drivers such as price and promotions | Need feature engineering and careful validation |
| Deep learning and time-series foundation models (for example Amazon Chronos, Google TimesFM, Salesforce Moirai) | Can forecast with little history for a series; pretrained on large collections of time series | Newer; published benchmarks are mostly the vendors' own; licences differ by version, and some recent model weights are licensed for non-commercial or research use only |
Pro tip
Large forecasting competitions have repeatedly found that combining models and using machine learning across many related series works well at scale, but simple methods remain hard to beat for some series. Always compare against a seasonal naive baseline on your own data.
How to evaluate forecasts honestly
Evaluation is where many forecasting projects go wrong. A model that looks accurate on the data it was trained on can fail on next month.
- Backtest on past periods the model hasn't seen, using rolling origins: train up to a date, forecast the next period, move forward and repeat. This mirrors how the forecast will be used.
- Use more than one accuracy measure. MAE and RMSE are in the units being forecast. Percentage errors such as MAPE are easy to explain but break down when actual values are zero or near zero, and penalise over- and under-forecasting differently. Scaled measures such as MASE compare the model with a naive baseline. Weighted measures such as WAPE (total absolute error divided by total actuals) suit portfolios where high-volume items matter more.
- Measure bias, not just error. A forecast that is consistently 10% high causes overstock even if its average error looks acceptable.
- Evaluate the ranges. Check that an 80% prediction interval contains the actual value roughly 80% of the time.
- Evaluate at the decision level. If buyers order by product and warehouse each week, measure accuracy there, not only at company level.
- Measure value, not only accuracy. Ultimately you care about stockouts, excess inventory, service levels or cash shortfalls.
Architecture
- Data ingestion from ERP, point of sale, ecommerce, CRM and external sources (calendars, prices, weather), into a warehouse or lakehouse; see data engineering for growing businesses
- A feature and history layer with cleaned demand, stockout flags, promotions and hierarchies, versioned so each forecast can be reproduced
- A training and backtesting pipeline that runs baselines and candidate models on a schedule and records accuracy by segment
- A forecast store holding every forecast version with its ranges, model version and inputs
- A planner workspace to review forecasts, see the drivers, make overrides with reasons, and compare scenarios
- Publishing of approved forecasts to ERP, replenishment and financial planning systems through APIs
- Monitoring of accuracy and bias against actuals, with alerts when performance degrades or data stops arriving
Planning a forecasting system?
ZSpace Labs builds forecasting models and AI analytics and the planner applications and integrations that put forecasts to work in your ERP and planning processes.
Planner overrides and scenarios
Planners know things the data doesn't: a new store opening, a competitor's promotion, a supplier problem. The software should let them override forecasts, require a reason, and track whether overrides improved or worsened accuracy over time. That record is one of the most useful outputs of a forecasting system, because it shows where human judgement adds value and where it adds noise.
Scenarios answer "what if": what happens to demand and inventory if we raise prices 5%, run a promotion, or lose a supplier for two weeks? Keep scenarios explicit and labelled, separate from the baseline forecast, so plans built on them are clearly understood. This is where forecasting begins to support decisions rather than just predict numbers.
Build, buy or extend
Planning suites and ERP vendors include forecasting modules, and cloud providers offer forecasting tools and models. These suit standard retail and supply chain patterns well. A custom build, or a custom layer on top of these, makes sense when your drivers are distinctive, data is spread across several systems, you need forecasts embedded in your own workflows or product, or the packaged tools can't be evaluated transparently on your data. Check vendor roadmaps too: Amazon Forecast, for example, closed to new customers in July 2024, with AWS pointing users to other tools. Keep your data and pipeline portable.
A phased build
An illustrative order of work:
| Phase | Scope | Exit criteria |
|---|---|---|
| 1. Decision and data audit | Which decisions the forecast serves, at what level and horizon; data sources and quality | Agreed target, level, horizon and success measures |
| 2. Baseline | Clean history, stockout flags, seasonal naive and statistical baselines, backtesting | Baseline accuracy measured by segment |
| 3. Improved models | Drivers and machine learning models, compared against baselines | Measurable improvement where it matters |
| 4. Planner workflow | Review, overrides with reasons, publishing to ERP | Planners use the forecast instead of spreadsheets |
| 5. Scenarios and monitoring | What-if scenarios, accuracy and bias monitoring, retraining | Forecast performance tracked monthly with owners |
Common mistakes
- Treating stockout periods as low demand
- Choosing a model on training accuracy instead of backtests
- Reporting only MAPE, which misleads for low-volume items
- Ignoring bias and prediction intervals
- Forecasting at company level when decisions are made per product and location
- No record of planner overrides, so nobody learns whether they help
- Promising that AI will make forecasts "accurate" rather than better and better understood
Conclusion
Good forecasting software starts with the decision it supports, cleans the history honestly, beats simple baselines in backtests, shows ranges and bias, and gives planners a clear way to adjust and explain. Decision intelligence follows naturally once forecasts are trusted. For the difference between predictive models and generative AI, see predictive analytics vs generative AI.
To test plans against different assumptions, see scenario planning and simulation software.
Want better forecasts?
Tell us what you forecast today, how decisions are made and where your data lives. ZSpace Labs will suggest a realistic first step. Discuss your forecasting project.
Common questions.
Software that predicts future values such as demand, sales, cash flow or staffing from historical data and known drivers like prices, promotions and holidays, using statistical and machine learning models. Good systems show a range of likely outcomes, track accuracy against actuals, and let planners review and adjust forecasts before they are used.
It depends on the data, the level of detail and how far ahead you forecast. No model is always accurate. The honest test is a backtest on past periods the model hasn't seen, compared with a simple baseline such as last year's value for the same week, measured at the level where decisions are made.
Forecasting estimates what is likely to happen, such as how many units will sell next month. Decision intelligence combines forecasts with constraints, rules and scenarios to recommend what to do, and records decisions and their outcomes so they can be improved.
Use more than one. MAE and RMSE are easy to interpret in units. MAPE is familiar but breaks down when actual values are zero or small. MASE compares the model with a naive baseline, and WAPE weights errors by volume. Track bias and the coverage of prediction intervals as well.
Test it. Foundation models can be useful when many series have little history, but results vary by dataset and published benchmarks are mostly the vendors' own. Compare them with simple baselines and machine learning models in your own backtests, and check the licence for commercial use.