Our Methods
Kailo is currently on the third generation of marathon prediction methods. All of our methods work by creating a "fitness fingerprint" of an athlete by analyzing their training logs, then using this to estimate marathon fitness. Our estimate is a range of marathon times, which provide stretch times to aim for and conservative times that should feel comfortable.
We found our production methods have about half the error of existing marathon predictors on the market. But if our predictions seem off to you, they might be! We are in the process of incorporating more information, such as running elevation and training altitude, which the method is currently ignoring. Learn more about our methods below.
RaceSPT (BM-0)
SPT stands for Sensor Pretrained Transformer, our newest race prediction model that combines deep learning with physiological sensor data. Unlike traditional models that rely solely on pace and distance, RaceSPT uses a transformer architecture to understand the complex relationships between your training history, physiological response, and race course characteristics.
RaceSPT predicts your race performance using three key inputs:
Context
Your historical training logs provide the foundation for understanding your current fitness level and training adaptations. The model analyzes patterns in your training volume, intensity distribution, and progression over time.
Effort
Heart rate and cadence data from recent efforts tell us how your body responds to sustained running intensity. This physiological fingerprint helps the model calibrate predictions to your individual cardiovascular and biomechanical characteristics.
Course
Race distance, elevation profile, and altitude allow RaceSPT to account for terrain difficulty and pacing strategy throughout the race. The model can predict split times and adjust for hills, altitude, and course-specific challenges.
Data Included
- Training history
- Activity type
- Heart rate
- Cadence
- Altitude
- Eleveation
- Distance
- Time
- GPS
- Temperature
- Power
- Race course
- Race effort
Note: RaceSPT is currently available in beta through our benchmark tool. As we gather more data and validate predictions, this model will become the primary prediction method across Kailo.
OG++
This is the current method serving all predictions on Kailo. It is very similar to the OG Method, with the main adjustment being the addition of elements for the course and race conditions (e.g., hot and hilly). (The estimated intervals shrunk by about half when compared to the OG Method.) The method works by calculating a "running fitness fingerprint" using runs logged with GPS data, then uses this fingerprint to estimate a marathon time on a flat course on a cool day. It assumes you are racing tomorrow, so don't be surprised if after a big effort your prediction gets worse--this can be a good thing! Expect to see your prediction improve as you recover and continue training.
Data Included
- Race course type (e.g., hilly or flat)
- Race day temperature
- Taper mileage and statistics
- Training mileage and statistics
Not Included
- Training temperature and conditions
- Elevation
- Altitude
- Heartrate
- Non-running activities
- Treadmill runs
If you have been training on hilly terrain or in tough weather conditions, we've noticed the method tends to underestimate your fitness. Please bear with us–we're fixing that with the next batch of methods!
OG
Our first "spreadsheet" predictor tested with a very limited number of athletes. One of the nice features that elevated our predictor over existing products was providing a range of values (best-case to worst-case). This range was quite large, because we didn't have a lot of data, but we noticed that for many high-performing athletes, the best-case prediction was more accurate than any existing race predictor.
Data Included
- Training mileage and statistics
- Taper mileage and statistics