{
  "schema": "telecom-italia-mi-benchmark-v1",
  "generated_utc": "2026-09-09T21:26:47Z",
  "dataset": {
    "doi": "doi:10.7910/DVN/EGZHFV",
    "license": "ODbL 1.0 (Open Database License), per the Dataverse terms of use",
    "n_files": 62,
    "total_rows": 319896289,
    "first_day": "2013-11-01",
    "last_day": "2014-01-01"
  },
  "kpi": "internet_traffic",
  "aggregate": "hourly, summed over country codes and the six 10-minute intervals",
  "features": "calendar/cyclic time features and lags of the target only",
  "lags": [
    1,
    2,
    3,
    6,
    12,
    24
  ],
  "horizon": 24,
  "test_size": 0.2,
  "models": [
    "ridge_linear",
    "gradient_boosting",
    "mlp"
  ],
  "cells": {
    "5161": {
      "level": "high",
      "cell_id": 5161,
      "selection": {
        "cell_id": 5161,
        "total": 12740060.3472951,
        "rule": "largest total internet_traffic"
      },
      "hours": 1488,
      "first_hour": "2013-10-31 23:00:00+00:00",
      "last_hour": "2014-01-01 22:00:00+00:00",
      "split": {
        "rule": "time ordered, first 80% train, last 20% test, after lag features",
        "n_train": 1171,
        "n_test": 293,
        "first_test_hour": "2013-12-20 18:00:00+00:00"
      },
      "models": {
        "ridge_linear": {
          "rmse": 1697.5042124621202,
          "mae": 1361.4689204707308,
          "mape": 82.56058831457032,
          "fit_and_eval_s": 0.14
        },
        "gradient_boosting": {
          "rmse": 2793.534288578107,
          "mae": 1996.4735019144832,
          "mape": 108.83798119287394,
          "fit_and_eval_s": 0.65
        },
        "mlp": {
          "rmse": 7653.086555595464,
          "mae": 5258.9303047430285,
          "mape": 349.74984117309594,
          "fit_and_eval_s": 0.63
        }
      },
      "baselines": {
        "naive_last_value": {
          "rmse": 2407.7013445613225,
          "mae": 1521.2766024355008,
          "mape": 30.175189332260942
        },
        "seasonal_naive_24h": {
          "rmse": 4163.094151352722,
          "mae": 2457.0991140232168,
          "mape": 67.17030163314803
        },
        "n_test": 293,
        "note": "Same hold-out rows as the models, after the 24 h seasonal lag is available."
      }
    },
    "3168": {
      "level": "mid",
      "cell_id": 3168,
      "selection": {
        "cell_id": 3168,
        "total": 277931.19392290426,
        "rule": "total nearest the median (277871.1)"
      },
      "hours": 1488,
      "first_hour": "2013-10-31 23:00:00+00:00",
      "last_hour": "2014-01-01 22:00:00+00:00",
      "split": {
        "rule": "time ordered, first 80% train, last 20% test, after lag features",
        "n_train": 1171,
        "n_test": 293,
        "first_test_hour": "2013-12-20 18:00:00+00:00"
      },
      "models": {
        "ridge_linear": {
          "rmse": 24.25319779431583,
          "mae": 20.0543897094986,
          "mape": 17.327855702190277,
          "fit_and_eval_s": 0.02
        },
        "gradient_boosting": {
          "rmse": 26.87433750926734,
          "mae": 21.15849863883614,
          "mape": 18.494746121598453,
          "fit_and_eval_s": 0.29
        },
        "mlp": {
          "rmse": 245.8020253360201,
          "mae": 140.05828711414097,
          "mape": 126.74954121058153,
          "fit_and_eval_s": 0.64
        }
      },
      "baselines": {
        "naive_last_value": {
          "rmse": 15.897839622988672,
          "mae": 12.192728617139597,
          "mape": 9.608586611015793
        },
        "seasonal_naive_24h": {
          "rmse": 43.81992727022976,
          "mae": 28.6145967801771,
          "mape": 20.789686063204247
        },
        "n_test": 293,
        "note": "Same hold-out rows as the models, after the 24 h seasonal lag is available."
      }
    },
    "9408": {
      "level": "low",
      "cell_id": 9408,
      "selection": {
        "cell_id": 9408,
        "total": 51230.32474746043,
        "rule": "total nearest the 10th percentile (51224.7)"
      },
      "hours": 1488,
      "first_hour": "2013-10-31 23:00:00+00:00",
      "last_hour": "2014-01-01 22:00:00+00:00",
      "split": {
        "rule": "time ordered, first 80% train, last 20% test, after lag features",
        "n_train": 1171,
        "n_test": 293,
        "first_test_hour": "2013-12-20 18:00:00+00:00"
      },
      "models": {
        "ridge_linear": {
          "rmse": 5.981668811746116,
          "mae": 4.593342909554738,
          "mape": 16.081119023819237,
          "fit_and_eval_s": 0.02
        },
        "gradient_boosting": {
          "rmse": 5.11784887251231,
          "mae": 3.7916675461041436,
          "mape": 12.275535126064039,
          "fit_and_eval_s": 0.3
        },
        "mlp": {
          "rmse": 79.8236375040409,
          "mae": 41.78492517103062,
          "mape": 161.6701451615756,
          "fit_and_eval_s": 0.65
        }
      },
      "baselines": {
        "naive_last_value": {
          "rmse": 4.413518949717169,
          "mae": 3.5077397275671225,
          "mape": 11.626525276058205
        },
        "seasonal_naive_24h": {
          "rmse": 7.810390042656788,
          "mae": 5.6692384067467065,
          "mape": 18.28721746437268
        },
        "n_test": 293,
        "note": "Same hold-out rows as the models, after the 24 h seasonal lag is available."
      }
    }
  },
  "provenance": {
    "host": "aimlstation",
    "python": "3.12.3",
    "pandas": "3.0.3",
    "numpy": "1.26.4",
    "sklearn": "1.8.0",
    "git_sha": "25b666a",
    "wall_clock_s": 193.9
  },
  "note": "Forecast accuracy on public telecom data; per-cell hold-out metrics with the naive baselines beside them. Not an operator deployment result."
}