Overview
The land-value change is computed as if one new station were built at the scenario's location. The calculation combines the difference between predictions of a trained LightGBM model with a value that depends on the distance from the scenario's location. The locations whose land-value change is computed are the points of 国土数値情報 L01 地価公示 (2026). Falling locations are shown as well as rising ones.
Which locations are computed
Only locations within 2.5 times the impact radius of the scenario's location are computed. The impact radius is fixed per scenario for preset scenarios, between 1.5 and 3.0 km, and is the value AI estimates for free-text scenarios. When AI returns no impact radius, 2.0 km is used.
The model's prediction difference
For each location, the model predicts the price with the current station layout and with one station added at the scenario's location. The station layout uses the stations of 国土数値情報 N02 (2024 edition). The model's inputs are the location's longitude and latitude, the distances to the three nearest stations, the number of stations within 1 km and within 2 km, the distance to Tokyo Station, and the land-use class (commercial, residential or industrial). The current prediction is subtracted from the after-scenario prediction and the result divided by the current prediction; this is the model's change rate.
The distance-based value
Inside the impact radius, it is cap × (1 − (distance ÷ impact radius) to the power of 1.5). Outside the impact radius, it is cap × 0.1 × (1 − (distance − impact radius) ÷ impact radius), which is 0 for locations at 2 times the impact radius or farther. Because this value is never below 0, falls come only from the model's change rate. The cap is fixed per scenario for preset scenarios, and is the value AI estimates for free-text scenarios.
Combination and cap
The change rate is the sum of the model's change rate × 0.4 and the distance-based value × 0.6. The change rate is clipped so that it does not exceed the cap in either direction, up or down. The caps of the preset scenarios are 0.15 to 0.25 and are not shown on screen. When AI returns no cap for a free-text scenario, 0.15 is used. The instructions given to AI ask for a cap estimated in the range 0.0 to 0.3, but the server uses the returned value as it is.
Locations that are not shown
Locations whose change rate has an absolute value below 0.001 (0.1%) are left out of the result.
Predicted land value
The predicted land value is the location's current land value multiplied by (1 + change rate), rounded to whole yen. A location's current land value is its price in 国土数値情報 L01 地価公示 (2026). The change rate is converted to a percentage and rounded to one decimal place. The locations in the result are sorted by change rate, largest first.