How Our Learning Algorithm Works: Science Behind Smart Intervals

"Change your filter every 3 months" is one-size-fits-all advice that rarely fits anyone perfectly. Some homes need changes every 6 weeks. Others can go 5 months. The difference? Usage patterns, environmental conditions, and system characteristics.

Filterly's learning algorithm analyzes these factors to calculate personalized replacement intervals. Here's how it works.

The Traditional Approach (And Why It Fails)

Most filter reminders use time-based triggers:

• Residential: Every 90 days
• With pets: Every 60 days
• Allergies present: Every 45 days

This approach has three fundamental problems:

1. Ignores actual usage: A home in Phoenix running AC 18 hours daily needs far more frequent changes than a Seattle home running it 2 hours daily—even if both have pets.

2. Static assumptions: Seasonal variation matters enormously. Your January usage isn't your July usage.

3. No learning: The 90-day interval is the same in year 1 as year 5, despite accumulated data showing actual filter life.

Our Approach: Multi-Factor Learning Model

Filterly combines multiple data streams to calculate filter life dynamically:

1. Runtime Tracking

The most important factor is how much air flows through the filter. We track:

• Daily runtime hours: Total HVAC operation time
• Fan-only vs cooling/heating: Different modes stress filters differently
• Seasonal patterns: Summer and winter peak usage
• Cumulative airflow: Estimated cubic feet of air filtered

A filter's capacity isn't measured in days—it's measured in air volume processed. A home running the system 6 hours daily processes roughly the same air in 90 days as a home running 2 hours daily processes in 270 days.

2. Environmental Factors

Static home characteristics that affect filter life:

Location-based:

• Regional air quality (EPA AQI data)
• Pollen levels (seasonal variation)
• Dust and particulate matter
• Humidity levels

Home-specific:

• Square footage (affects system size and airflow)
• Number of occupants
• Pets (quantity and type—dogs shed more than cats, long-hair more than short-hair)
• Smoking in the home
• Construction or renovation activity

HVAC system:

• System age and efficiency
• Filter type and MERV rating
• Duct condition and seal quality

3. Historical Performance Data

As the system gathers data, it improves predictions:

Filter change feedback: When customers report changing filters, we validate our predictions. Did we recommend it too early? Too late? Adjust the model.

Seasonal refinement: After one year, we have complete seasonal data. Year 2 predictions are far more accurate than year 1.

System degradation: HVAC systems become less efficient over time. We factor this into runtime calculations.

The Algorithm: Step by Step

Here's the actual calculation process:

Step 1: Establish Baseline Capacity

Each filter type has a rated capacity based on MERV rating and size. For example:

• 16x25x1 MERV 8: ~45,000 cubic feet capacity
• 16x25x1 MERV 13: ~35,000 cubic feet capacity (higher filtration = faster saturation)

This is the starting point, but real-world capacity varies based on installation quality and system characteristics.

Step 2: Calculate Daily Airflow

Estimate air volume processed daily:

Daily Airflow = (Runtime Hours) × (System CFM) × 60 minutes

System CFM (cubic feet per minute) is estimated from:

• Home square footage
• Typical HVAC sizing rules (400-450 CFM per ton)
• Reported system capacity if available

A 2,000 sq ft home typically has a 3-4 ton system (~1,200-1,600 CFM). If it runs 6 hours daily:

6 hours × 1,400 CFM × 60 = 504,000 cubic feet/day

Step 3: Apply Environmental Multipliers

Adjust baseline capacity based on conditions:

Pets:

• 1 dog/cat: 0.85× capacity (filters clog 15% faster)
• 2+ dogs/cats: 0.70× capacity
• Long-hair breeds: Additional 10% reduction

Air quality:

• EPA AQI < 50 (good): 1.0× capacity
• AQI 50-100 (moderate): 0.90× capacity
• AQI 100+ (unhealthy): 0.75× capacity

Seasonal factors:

• High pollen season: 0.85× capacity
• Construction nearby: 0.70× capacity
• Wildfire smoke events: 0.60× capacity

Step 4: Calculate Projected Filter Life

Remaining Days = (Remaining Capacity) / (Daily Airflow)

Example calculation:

• Filter: 16x25x1 MERV 11, 40,000 cubic feet capacity
• Home: 2 dogs, Phoenix summer (high AC usage)
• Adjusted capacity: 40,000 × 0.75 (pets) × 0.90 (AQI) = 27,000 cubic feet
• Daily airflow: 720,000 cubic feet (18 hours × 1,400 CFM × 60)
• Filter life: 27,000 / 720,000 = 38 days

Same filter in Seattle winter with no pets:

• Adjusted capacity: 40,000 × 1.0 × 1.0 = 40,000 cubic feet
• Daily airflow: 168,000 cubic feet (2 hours × 1,400 CFM × 60)
• Filter life: 40,000 / 168,000 = 238 days

Massive difference—yet both homes would get "90 days" with traditional reminders.

Step 5: Continuous Refinement

After each filter change, we compare predicted vs actual life:

Adjustment Factor = (Actual Days) / (Predicted Days)

If we predicted 45 days but the customer changed at 60 days (filter still had life), we adjust the model to be less aggressive. If they changed at 30 days (filter was fully saturated), we adjust to be more conservative.

Over 3-4 cycles, the model converges on highly accurate predictions for that specific home.

Machine Learning Layer

On top of the physics-based calculations, we apply machine learning to identify patterns:

Clustering homes with similar profiles: Homes with comparable characteristics (location, size, pets, usage) likely have similar filter life. We use this to improve predictions for new users.

Seasonal pattern recognition: ML identifies usage spikes and dips better than rule-based logic. It learns that Phoenix sees extreme summer usage, Seattle sees dual winter/summer peaks, and Florida runs year-round.

Anomaly detection: Sudden changes in runtime might indicate issues (stuck relay, thermostat malfunction). We flag these for customer attention.

Privacy and Data Usage

All calculations run on aggregated, anonymized data. Individual home data is never shared. The ML model learns from population patterns but applies them to individual predictions.

We use:

• Aggregated runtime statistics (not individual timestamps)
• Regional environmental data (not home addresses)
• Filter change frequencies (not customer identities)

Real-World Accuracy

After 12 months of learning, the algorithm typically achieves:

• ±7 days accuracy for 90% of predictions
• ±3 days accuracy for 70% of predictions
• Zero premature recommendations (filter still has 20%+ life) in 95% of cases

This precision means customers aren't wasting money on early changes or risking system issues from late changes.

Why This Matters

Accurate filter life prediction delivers:

Customer savings: No wasted filters replaced prematurely System protection: Filters changed before airflow restriction damages equipment Energy efficiency: Clean filters maintain optimal airflow and efficiency Trust: Customers see that recommendations match their actual experience

The Future: IoT Integration

The next frontier is direct system integration. Smart thermostats and sensors can provide:

• Real-time runtime data (no estimation needed)
• Airflow pressure monitoring (detects clogged filters)
• Indoor air quality sensors (triggers early changes if needed)

We're building partnerships to integrate with major smart home platforms. The goal: zero-friction filter tracking that requires no customer input.

Conclusion

"Every 3 months" was always an educated guess. Now, data-driven algorithms can calculate actual filter life with precision.

The result isn't just better reminders—it's HVAC systems that run more efficiently, last longer, and require less reactive maintenance. All because the filter gets changed at exactly the right time.

For HVAC professionals: This technology is available now. Your customers expect this level of service.

For homeowners: Demand filter tracking that adapts to your home, not generic advice.

The science is proven. The results are measurable. Smart filter intervals aren't the future—they're the present.