DataByte

Forecast what happens next, and how sure you are

Forecaster puts 25+ time-series algorithms under one governed surface. Scheduled runs keep forecasts current, side-by-side backtests decide which model earns the job, and every prediction ships with the confidence range that tells you how much weight it can carry.

Forecast horizons and confidence bandsCash Forecaster pairs a model with each horizon: LightGBM for the short term, XGBoost with ARIMA for the medium term, and Prophet for the long term. Below, a chart shows actuals to date and a forecast whose P10 to P90 confidence band widens further out.FORECAST HORIZONSShort termLightGBMQuantile regression tuned fordaily and end-of-month patternsMedium termXGBoost + ARIMARolling 13-weekprojectionLong termProphetMacroeconomic factors,out to twelve monthsTODAYP90P10P10–P90 confidence bandForecastActuals
See it running

Empowering Businesses with Predictive Insights | Forecaster

A walkthrough of Forecaster, DataByte’s enterprise time-series forecasting agent with automated preparation, 25+ algorithms, AutoML, and scheduling.

Algorithms

Twenty-five algorithms, and a comparison you can defend

Different series want different models. A daily operational metric with a hard weekly rhythm is not the same problem as a twelve-month financial trend, and picking one algorithm for both is how forecasts quietly go wrong.

Forecaster ships more than twenty-five time-series algorithms in a single library: classical statistical models such as ARIMA and its seasonal extension SARIMA, gradient-boosted learners such as LightGBM and XGBoost, Prophet, and deep-learning approaches such as N-BEATS.

Rather than settle the argument in a meeting, settle it with a backtest. Forecaster runs candidate models against the same history and reports the results side by side, so the choice is made on evidence.

Accuracy dashboards then keep reporting MAE, the typical error in the units you actually care about, and MAPE, the average percentage miss. Scheduled runs keep the winning model producing on a cadence. Nobody has to remember to refresh a spreadsheet.

Forecaster algorithm library and backtestingSix named algorithms out of more than twenty-five in the Forecaster library feed a side-by-side backtest, which in turn feeds accuracy dashboards reporting MAE and MAPE and the scheduled runs that keep forecasts current.25+ TIME-SERIES ALGORITHMSARIMASARIMAProphetLightGBMXGBoostN-BEATSStatistical, gradient-boosted, and deep-learning families in one librarySide-by-side backtestCandidate models run against the same history, then comparedAccuracy dashboardsMAE, the typical error in unitsMAPE, the average percentage errorScheduled runsForecasts refresh on a schedule,not when someone remembers
Data readiness

Most forecasting failures are data failures in disguise

Cash Forecaster, the treasury configuration of the module, makes the point concretely. It draws on three live feeds: bank statements in BAI2 and MT940 format, ERP extracts covering receivables and payables, and the payroll calendar.

Each arrives in its own shape, on its own schedule, with its own idea of what counts as an event. A model handed that directly will produce a number, and the number will be worthless.

So five stages sit between the feeds and the model: raw collection, normalisation that deduplicates records and standardises event types, entity-level reconciliation into a unified ledger, feature engineering that derives signals like seasonality indicators and rolling averages, and a final completeness check before anything reaches inference.

An input coverage score aggregates how much of that arrived intact, across bank statement coverage, ledger completeness, payroll coverage, feature derivation, and historical backfill depth.

Drop below eighty percent and the platform pauses rather than publish. It would rather show you nothing than show you something wrong.

Cash Forecaster five-stage normalisation pipelineBank feeds in BAI2 and MT940 format, ERP receivable and payable extracts, and the payroll calendar pass through five stages: raw collection, normalisation, entity-level reconciliation, feature engineering, and a validated model input record. An input coverage score gates the whole pipeline.FROM RAW FEEDS TO A TRUSTED FORECASTBank feedsBAI2 and MT940ERP extractAR aging and AP ledgerPayroll calendarScheduled obligations1Raw sourcesCollect every feed2NormaliseDeduplicate, standardise event types3Unified ledgerEntity-level reconciliation4Feature engineeringSeasonality, rolling averages5Model input recordValidated, inference-readyINPUT COVERAGE SCOREBelow 80 percent, the platform pauses inference rather than publish a forecast
Scenario analysis

What happens if the money arrives ten days late

Every planning conversation eventually turns into a what-if. What if our largest customer stretches payment? What if revenue softens?

What if we accelerate payments to protect a critical supplier? Those questions are usually answered with instinct and rough arithmetic, which is fine right up until the answer matters.

The scenario module turns them into structured analysis. You start from the live base case, the same model driving the main dashboard, then adjust bounded parameters: DSO extension, collection acceleration, AP payment stretch, and revenue growth.

Stress, base case, and upside run on the same underlying model, so a stress test is a real projection rather than a hypothetical drawn on a whiteboard.

Results overlay as P25, P50, and P75 on the fourteen-day chart, with a summary table that compares all three at once and shows each against your minimum liquidity threshold.

Scenario analysis frameworkThree columns, stress, base case, and upside, each listing the parameters it adjusts: days sales outstanding, collection acceleration, revenue growth, and accounts payable stretch. Every scenario runs on top of the same live base-case model.SCENARIO ANALYSISStressIf things go badly, are we safe?PARAMETERSDSO extension, longerCollection acceleration, lowerRevenue growth, lowerAP stretch, pay on timeBase caseWhere is cash heading as things stand?PARAMETERSDSO, current actualsCollections, learned patternsRevenue, trailing trendAP, scheduled paymentsUpsideIs there surplus to put to work?PARAMETERSDSO extension, shorterCollection acceleration, higherRevenue growth, higherAP stretch, longerONE LIVE MODEL UNDERNEATH ALL THREEResults overlay as P25, P50, and P75 on the 14-day forecast chart, side by side
Model health

A forecast that tells you when it is starting to be wrong

Models decay. They learn historical patterns, and when the business moves, those patterns stop holding. The damage is rarely dramatic; the forecast simply drifts, quietly, for weeks. Forecaster treats that as a monitoring problem rather than an act of faith.

Every horizon carries its own model card showing the algorithm in use, current MAPE against its threshold, an eight-week accuracy trend, the last retrain date, and the training data range.

Two drift signals run underneath. Population Stability Index watches whether the distribution of an input feature has shifted since training, and forecast error drift watches whether predictions are becoming systematically biased in one direction, the failure mode an acceptable average error will happily hide.

SHAP values rank feature importance at each horizon, so the model can explain which variables are driving it. When a MAPE threshold is breached, retraining triggers automatically, with optional AutoML tuning on the run.

The confidence score on the dashboard composes all of it: historical accuracy at forty percent, horizon decay at twenty-five, data freshness at twenty, and forecast stability at fifteen.

Forecast model health monitoringThree monitors, feature drift measured with the Population Stability Index, forecast error drift, and SHAP feature importance, feed a MAPE threshold check per horizon. Breaching the threshold triggers retraining with AutoML tuning. A bar shows the confidence score composition: historical accuracy 40 percent, horizon decay 25 percent, data freshness 20 percent, forecast stability 15 percent.WATCHING THE MODELFeature driftPopulation Stability IndexHas the input distributionmoved since training?Error driftSystematic bias checkAre errors leaning the sameway week after week?Feature importanceSHAP valuesWhich variables drivepredictions at each horizon?MAPE against its threshold, per horizonEach horizon carries its own model card, trend, and last retrain dateThreshold breached, retrainOptional AutoML tuning optimises hyperparameters on the retrain runCONFIDENCE SCORE COMPOSITIONHistorical accuracy40%Horizon decay25%Data freshness20%Forecast stability15%

Put a number, and a range, on next quarter.

See Forecaster backtest candidate models on your own history, schedule the winner, and report accuracy as it runs.