Analytics

The IoT analytics layer that predicts.

Dashboards tell you what already happened. TagoIO analytics reads the telemetry you already collect and tells you what happens next: forecasts with confidence bands, anomaly scores against learned normal behavior, and countdowns to the moment a variable crosses a limit. No ML code and no data science team required.

Create your Model Get Started

Business impact

What would predictions save in your operation?

From reacting to predicting, in dollars per year

Pick a sector, set your numbers, and see the yearly impact of moving from reacting to predicting. Directional estimates with adjustable assumptions, not a quote.

Reacting
$120K/yr
Predicting
$24K/yr

$96K saved per year

384 emergency deliveries become scheduled routes each year. Level forecasts order grain days before the bin runs empty.

Assumption: with days-ahead level forecasts, 80% of emergency refills become scheduled deliveries. Savings = bins × emergencies × premium × 80%. See the smart feed bin use case for a production example. Powered by the Forecasting capability below.

Capabilities

What can IoT analytics do with your sensor data?

Three models, one wizard, zero ML code

Three capabilities, each trained on the history already in your device buckets, each writing results back as regular variables. Setup is a guided wizard: pick devices, pick signals, choose a behavior, install.

Anomaly Detection

Catch problems before they reach your dashboard. Score every device against a learned baseline of normal behavior. Flag drift, mode shifts, and peer outliers.

See how

Forecasting

See what your telemetry is about to do. Predict future values with confidence bands. Factor in drivers like weather and load.

See how

Time to Threshold

Know how long you have, not just what's coming. Forecast a variable and count down to the moment it crosses a limit, with multi-tier severity.

See how

Your telemetry

Device buckets, months of history at any resolution

Trained model

Forecasting, anomaly, time to threshold, on your cadence

Variables

Forecasts, scores, ETAs, confidence bands, as plain TagoIO variables

Your application

Dashboards, Actions and alerts, TagoRUN portals

Models train on data already in TagoIO and write predictions back as regular variables. Nothing downstream changes.

Interactive

See it work.

Live simulator with synthetic telemetry

Synthetic telemetry, real behavior. Switch capability and adjust the controls.

now

Output: _ts_forecast for the next 48 hours, each point with upper and lower confidence bounds. Note how the band widens: the model is honest about distance.

In your account, these outputs are variables: forecasts with upper and lower bounds, anomaly scores and flags, anomaly windows, ETAs, and severity tiers. Alarms, dashboards, and integrations use them the way they use any sensor reading.

Forecasting

From history to what happens next.

MSTL + AutoETS · Prophet · OLS regression

Sensor data breathes in daily and weekly cycles, and TagoIO forecasting is built around that fact.

Forecast values. The engine detects your data's cycle, forecasts with an automatically selected seasonal model (MSTL + AutoETS), and layers the cycle back over the horizon. No cycle means an honest drift-following forecast, not an invented oscillation.

Demand prediction. Forecast a target from drivers like weather or load: each driver is forecast forward, then feeds a Prophet model for the target.

Estimate from drivers. A regression estimates a target from variables you already measure, for gap-filling and sensor sanity checks. It explains the present; it does not forecast.

Forecast with confidence band

now
The band widens with the horizon: honest uncertainty.

Anomaly score with flagged windows

flag threshold 0.60
Outputs: a score, a flag you can alarm on, and windows that shade the whole span.

Anomaly Detection

Four questions, four models.

Isolation Forest · K-Means · DBSCAN · EWMA

"Is something wrong?" is really four questions, each with the model designed for it.

Is this reading unusual? (Quick health check.) An Isolation Forest scores each reading against the joint picture of normal, catching combinations that are individually in range but collectively wrong.

Which operating mode is this? (Continuous monitoring.) K-Means learns your fleet's normal operating modes and scores each reading by distance from the nearest one.

Which device is the odd one out? (Peer comparison.) Density clustering (DBSCAN) groups devices that behave alike and flags the one that belongs to no group.

Is something slowly going wrong? (Slow drift.) An EWMA tracked against your seasonal baseline catches the slow creep that defeats fixed thresholds.

Time to Threshold

A countdown, not a surprise.

ETA and severity from one configuration

Alerts fire when a limit is crossed; by then the tank is empty. Time to Threshold forecasts the variable, scans for the crossing, and publishes an ETA plus a severity tier you configure once.

Time to Threshold: the countdown

now threshold 24 ETA: 49h
ETA and severity land as variables, ready for gauges and escalation Actions.

Integration

From prediction to action.

Variables in, everything downstream unchanged

Predictions land as regular variables, and everything downstream already knows what to do with variables:

Dashboards

Chart forecasts next to live readings and shade anomaly windows, fleet-wide via Blueprints.

Actions

Trigger on predictions like on sensor values: "days-to-empty below 4" opens a resupply order.

TagoRUN

Portals expose predictions to your customers under your brand.

Analysis

Scripts in Node.js, Deno, or Python post-process outputs when you need custom logic on top.

Outcomes

What predictions are worth in your operation.

Five sectors, one mechanism

See problems while they are still cheap:

Agriculture

Feed-bin levels forecast days ahead; grain ordered on forecast, not emergency.

See the use case

Energy

Consumption anomalies flagged the day they start, not at month's end.

Buildings

HVAC drift caught while it is a maintenance visit, not a tenant complaint.

Industrial

"The pump failed" becomes "here is the window to service it."

Logistics & cold chain

Hours of warning before a reefer breaches compliance.

Predictions convert unplanned events into scheduled work, and scheduled work is always cheaper.

Under the hood

Built on published science, engineered for telemetry.

Known methods, documented failure modes

Every model here is a published, peer-reviewed method:

Capability Method
Seasonal forecasting MSTL decomposition + ETS exponential smoothing
Demand prediction Prophet with driver forecasts
Estimation from drivers Multivariate OLS regression
Quick health check Isolation Forest
Continuous monitoring K-Means clustering
Peer comparison DBSCAN density clustering
Slow drift EWMA control chart on a seasonal baseline
Time to Threshold Holt-Winters / AutoETS forecast + threshold scan

Our work is the part no paper covers: making these methods behave unattended, across a fleet, on real telemetry.

Frequently asked questions

Quick answers about TagoIO analytics

What is IoT analytics?
IoT analytics is the processing of sensor and device telemetry to produce insight beyond raw readings: forecasts of future values, detection of anomalous behavior, and predictions such as time until a threshold is crossed. On TagoIO, analytics models train on your existing device data and return results as regular variables for dashboards, alerts, and applications.
Can TagoIO predict equipment failure?
TagoIO detects the signatures that precede failure: anomaly scores against learned normal behavior, drift detection for slow degradation, peer comparison to flag the device behaving unlike its siblings, and time-to-threshold countdowns for variables trending toward limits. These deliver a service window before the breakdown without requiring labeled failure history.
Do I need a data science team to use TagoIO analytics?
No. Models are configured through a guided wizard: pick devices, pick signals, choose how the model should behave, and install. Sensible defaults come from the data itself. Developers who want to go further can post-process outputs with Analysis scripts in Node.js, Deno, or Python.
Which forecasting models does TagoIO use?
Seasonal and trend forecasting uses MSTL decomposition with automatically selected ETS (exponential smoothing) models; multivariate demand prediction uses Prophet fed by driver forecasts; estimation uses multivariate OLS regression; and time-to-threshold uses Holt-Winters or AutoETS forecasts scanned for the crossing. Model selection is automatic based on your data's cycles and history.
How do predictions appear in my application?
As regular TagoIO variables: forecast values with upper and lower confidence bounds, anomaly scores and flags, anomaly windows with start and end dates, ETAs in hours, and severity tiers. Dashboards, Actions, the API, and TagoRUN portals consume them like any sensor reading.
How much history do I need to train a model?
As a rule, train on 3 to 4 times more history than the range you want to predict; forecasting one month ahead wants three to four months of data. Seasonal models need at least two full cycles of the pattern they learn. The setup wizard checks this and tells you when your window is short.

Get started

Connect. Predict. Act.

Set up your first forecast in minutes. The free plan includes 5 devices and 5 dashboards.

Create your Model Get Started