← Back to Blog

How ThresholdIQ Detects Anomalies Automatically (No Setup Required)

Published March 14, 2026 · 9 min read · Tutorial

Most monitoring tools ask you to do the hard work first: decide which columns to watch, pick your thresholds, and configure your rules. But here's the problem — if you already knew exactly what was wrong and where to look, you wouldn't need a monitoring tool.

ThresholdIQ takes the opposite approach. Upload your file. Click detect. The engine figures out what's unusual. This post explains exactly how that works, in plain English.

The Problem With Manual Thresholds

Imagine you're a Finance Analyst monitoring daily revenue. You set a rule: "alert me if revenue drops below £50,000." Sensible. But:

A static threshold catches the obvious cases. It misses the subtle, the seasonal, and the correlated ones. ThresholdIQ's automatic engine is designed to catch all of them.

What Happens When You Click "Detect Anomalies"

When you upload a file and click the button, nine detection methods run simultaneously across every numeric column in your data. Here's what each one does and what it catches:

Method What it does What it catches Severity
Multi-Window Z-score Computes local mean and standard deviation over 50, 100, 200 and 500-point rolling windows. Flags values that deviate more than 2–3.5 standard deviations from recent history. Sudden spikes, abrupt drops, sustained departures from normal range Warning → Emergency
EWMA Applies exponential weighting so recent values count more than older ones. Flags residuals between the raw value and the smoothed trend. Fast, sharp spikes that would be missed by a slow rolling average Boosts score
SARIMA Seasonal ARIMA model learns the regular cycle in your data (e.g. daily/weekly patterns). Flags points that deviate from seasonal expectation, not just raw magnitude. Anomalies that look normal in isolation but are wrong for that time of day or day of week Boosts score
Isolation Forest Treats every row as a point in multi-dimensional space (one dimension per metric). Identifies points that are isolated — far from all other points — across all metrics at once. Global outliers, sensor failures, zero readings when value should be non-zero Emergency
Correlation Deviation Monitors whether correlated metrics deviate together in the same direction. Two or more metrics all abnormal simultaneously is a stronger signal than one alone. Multi-metric failures — e.g. revenue AND margin AND volume all drop together Emergency
DBSCAN Groups your data points into "normal behaviour clusters." Points that don't belong to any cluster are labelled noise and flagged. Behavioural outliers — patterns that don't match any known operating mode Critical
Seasonal Baseline Builds a separate mean and standard deviation for every hour-of-day and day-of-week bucket. Sunday overnight readings are compared against Sunday overnight history — not all-time history. Prevents false alarms from normal seasonal lows; surfaces genuine anomalies within their time context Warning
Trend Detection Compares the average of three consecutive 50-point windows. A monotonic rising or falling drift across all three windows is flagged. Gradual budget drift, slow inventory decline, creeping latency — things that look fine today but signal a problem forming Critical
Stuck/Zero Detection Detects when a series that previously had variation becomes constant, or drops to zero from a meaningful non-zero history. Sensor failures, data pipeline outages, broken integrations that produce zeroes instead of real values Emergency

How the Results Are Combined

Each method produces a score between 0 and 1 for every data point. These scores are combined using a weighted fusion formula:

Final score = Multi-Window score + min(0.25, ML composite × 0.25)

The Multi-Window Z-score is the primary driver of severity. The other eight methods can only boost a score — they can never reduce it. This prevents a single false-positive method from masking a real anomaly.

The final score maps to a severity level:

A Real Example: IoT Sensor Data

Suppose you upload a CSV of hourly temperature readings from three factory facilities over 10 days. Here's what happens:

1
Schema detection: ThresholdIQ finds the timestamp column, the three numeric temperature columns, and the facility dimension. No mapping required.
2
Seasonal baseline: The engine notices temperature is reliably lower overnight (hours 22–06). It builds a separate baseline for each hour slot.
3
Multi-window scoring: A spike at 03:00 Tuesday gets a high W50 score. But the seasonal baseline confirms this is within normal overnight range — so the SARIMA method returns a low residual. Final score stays at Warning.
4
Stuck detection: Facility-C's temperature reads exactly 21.4°C for 20 consecutive hours on Thursday. Stuck detection fires — score jumps to Emergency. This is a sensor failure, not a real reading.
5
Results: The timeline shows Facility-C flagged in red from Thursday 08:00. The Detection Signals tab shows "Stuck/Zero: 1" — exactly one method fired, and it fired confidently.

How Much Data Do You Need?

Data minimums:

For most Finance or Operations exports — weekly KPI reports, monthly actuals, daily ops logs — you'll have well over 100 rows and the full engine runs immediately.

What About False Positives?

This is the most common concern with automated detection. ThresholdIQ addresses it in three ways:

  1. Seasonal awareness: SARIMA and the hourly/daily seasonal baseline prevent routine low periods from triggering alerts.
  2. Multi-window confirmation: A Warning only becomes Critical when it persists into the 100-point window. Transient spikes stay at Warning level.
  3. Fusion capping: ML methods can add at most 0.25 to the base score. A single method can't manufacture a Critical alert on its own.

What You See in the App

After detection completes, ThresholdIQ gives you four views:

Try It Free — Upload Your File and See What the Engine Finds →

No thresholds to configure. No rules to write. Just upload and detect.