Start with a second reading you can compare
A Home Assistant filter sensor is useful when one numeric reading jumps around and you want a steadier value. Start with a modest Lowpass filter for everyday room-temperature display. This blends each incoming value with the previous filtered result. It does not calibrate the thermometer or establish which reading is true.
First decide whether you want to change the number or only its appearance. A dashboard’s averaged graph can look calmer while your automation still receives the original sensor. Conversely, creating a Filter helper does not redirect existing automations: they keep using whichever entity you selected. An entity is Home Assistant’s name for one reading or control.
Our recommendation is to compare the raw and filtered readings on a dashboard before using the new value elsewhere. If a fan or heater keeps switching around one boundary, first consider separate on and off thresholds. Filtering the input can add delay without fixing the switching rule. The Threshold helper calls that separation hysteresis.
Researched from current Home Assistant documentation and Core 2026.10.2 source on October 22, 2026. The numerical example is calculated, not measured in a Tara installation. No physical sensor or heating system was tested. The cover is an AI-generated residential illustration.
1. Choose the smallest change that solves the problem
Look at the original sensor’s recent history. Note whether you see frequent small changes, occasional implausible jumps, or long gaps. Check its placement and manufacturer’s instructions before assuming the data needs filtering.
On a narrow screen, scroll the table sideways to read every column.
| What you need | Starting point | What to watch |
|---|---|---|
| A calmer chart only | A Statistics graph card showing the mean, if the sensor provides suitable statistics. | This changes the display, not the sensor used by a routine. Keep minimum and maximum values available when peaks matter. |
| A steadier value from frequent small fluctuations | Filter → Lowpass. | Real changes also take longer to appear fully. Compare the result with the original. |
| An occasional implausible jump | Investigate the sensor; consider Outlier only after you know its normal behavior. | Outlier substitutes the recent median outside a chosen band. A real sudden change may be rejected too. |
| A device switching repeatedly near one temperature | Separate on/off thresholds, with equipment-appropriate control rules. | Hysteresis is not a minimum run-time or compressor-protection feature. Keep the equipment’s own protections. |
| One current reading for several rooms | A Sensor group using the selected rooms. | This combines different sensors; it does not smooth one sensor over time. |
The Statistics graph documentation explains display averages. For the last row, use our guide to combining room temperatures. If you need a rolling statistic over a defined period, the separate Statistics helper is another option; its time window and sample capacity both need attention.
2. Add a Lowpass helper beside the original sensor
You need a working numeric sensor entity and an account that can create helpers. No extra device, paid subscription, app or custom integration is required for this calculation. The original device’s connection still has its own requirements.
- Open Settings → Devices & services → Helpers → Create helper → Filter.
- Name it something explicit, such as Living room temperature smoothed. Select the original room-temperature sensor as the input.
- Choose Lowpass. For this demonstration, set Time constant to 4, Precision to 2, and leave Window size at 1. These are learning settings, not a universal thermostat recommendation.
- Save. Find the new entity and add it beside the original reading on a dashboard. A new helper may need recorded starting data or a normal source update before it has a value. Once both readings are available, confirm they show the same temperature unit.
- Leave existing automations on their original inputs while you compare the two. Record the exact new entity ID if you later choose to use it.
The current Filter documentation describes the helper route. The checked 2026.10.2 configuration flow exposes these Lowpass settings and accepts sensor entities. Older versions may use different labels. If Filter is missing from your helper list, check your installed version before following an older YAML-only tutorial.
Time constant 4 does not mean four seconds. At each processed input, this example keeps three quarters of the previous filtered value and adds one quarter of the new reading. A larger constant gives the newest reading less influence. Two decimal places make the demonstration easier to follow; they do not make the physical sensor more accurate.
3. Read a worked example before trusting the result
Imagine the raw and filtered temperatures have already settled at 20.00°C. The following source changes arrive in order, with no extra updates between them. With Time constant 4 and Precision 2, the calculation is:
New filtered value = (previous filtered value × 0.75) + (new raw value × 0.25), rounded to two decimal places for each step.
These are calculated values, not hardware measurements. Scroll sideways on a narrow screen.
| Incoming event | Raw reading | Filtered result | Meaning |
|---|---|---|---|
| Starting assumption | 20.00°C | 20.00°C | Both readings are already settled. |
| One sudden jump | 24.00°C | 21.00°C | The peak is reduced, not removed. |
| Return to the earlier reading | 20.00°C | 20.75°C | The earlier peak still influences the result. |
| A later change | 21.00°C | 20.81°C | The output now trails the raw reading. |
| No further source state-change event | Still displayed as 21.00°C | Still 20.81°C | Elapsed time alone does not finish the calculation. |
The last row matters for sensors that report only changes. The built-in filter listens to source state-change events; it is not continuously moving toward the last number on its own clock. An attribute change can also produce a source state-change event. Actual results depend on the events your integration supplies, as well as the starting history. The Filter implementation defines that update path and calculation.
Do not inject random numbers into a measurement to make it look active. If a stable input must be represented directly, keep the raw value for that purpose. If you need a time-based statistic, choose and verify that calculation explicitly instead of treating Lowpass as a timer.
4. Verify it without switching equipment
- Keep this a display-only check. Open both entities’ history and current values. Do not change a thermostat target, heat a freezer sensor, or edit a real sensor’s state to create a test.
- Observe several normal updates. Expect the filtered reading to move less abruptly than the raw one. After a real sustained change, expect some lag. Immediately after creation or restart, starting history can affect the comparison.
- Check one calculation. Note the previous filtered value and next raw value, then use the formula above. If your result differs, inspect intervening state or attribute updates before changing settings.
- Check a quiet period. If there are no source state-change events, a Lowpass output can stay behind the displayed raw value. Decide whether that behavior is acceptable for your intended use.
- Keep a clear way back. Leaving the helper unused by automations makes this trial reversible. If you later select it in a routine, document the original entity so you can restore that input.
There is no need to restart the house or delete recorded history for this comparison. A smoother line alone is not a pass: the result must still represent changes quickly enough for the job. If the difference is hard to judge, continue using the original sensor.
If the reading looks wrong
Scroll sideways on a narrow screen.
| Observation | Check first | Useful next step |
|---|---|---|
| The smoothed value barely moves | Source update cadence and Time constant. | Compare real incoming events. Try less smoothing for a display, or keep the raw input where responsiveness matters. |
| A large spike still appears | Whether Lowpass is the only selected filter. | Lowpass softens a spike; it does not identify errors. Investigate the source before choosing an Outlier band. |
| The source is unavailable | The original entity and its device connection. | Fix that path. A filter cannot supply a missing measurement. Keep an explicit unavailable-device warning if it matters. |
| The number looks current but never changes | The source’s own reporting and availability behavior. | A frozen numeric value is not proof of fresh data. Compare the device’s expected reporting before diagnosing failure. |
| An existing automation behaves exactly as before | Which entity its trigger and conditions use. | That is expected if they still reference the raw sensor. Change them only after the comparison supports it. |
For connection warnings, see our unavailable-device notification guide. It distinguishes an unavailable state from an unchanged value. Filtering is not a substitute for that distinction.
Common questions
Should I add Outlier and Lowpass together?
Only if you have identified both isolated bad samples and ordinary fluctuations. The UI configures one filter per helper; YAML can apply a chain in its listed order. Start with one so you can see what it changes. Outlier needs enough previous samples for its comparison, and a real abrupt change can resemble an error. Do not copy somebody else’s radius without checking the unit and normal range.
Will this reduce my database size?
Creating a second entity is not a database cleanup. The original sensor can still be recorded, as the Recorder reference explains. Throttle filters skip samples; they do not average them or automatically exclude the original from Recorder. Keep storage changes separate from this display trial.
Can I use the smoothed value for heating or an urgent alert?
Do not move an urgent alert onto a slower or spike-rejecting input just to reduce noise. Heating also needs appropriate equipment controls, failure behavior and cycling limits. This guide verifies a calculated display; it does not validate a heating controller or replace a certified alarm.
The filter reference covers chaining and sample selection; the versioned implementation shows Outlier’s initial sample requirement. For a household plan, Tara’s home automation kit with coordinated local routines is the relevant next step when you want the inputs, exceptions and everyday controls documented together.