25 automatic changepoints

match ← All linear growth benchmarks

256 daily rows with weekly seasonality and Prophet's 25 automatic changepoints, so the posterior carries real uncertainty about trend changes into 64 future rows. mcmc_samples=300.

Input

benchmark/inputs/v1/linear-medium.json: 256 history rows and 64 forecast rows.

Output

Largest difference: 0.00088 (0.00091% of the input's range).

Every forecast row from both libraries.

Date	Effect Prophet	Python Prophet	Difference
2020-09-13	107.7963	107.7970	0.00069
2020-09-14	107.5643	107.5648	0.00055
2020-09-15	108.3430	108.3430	1.1e-6
2020-09-16	109.9339	109.9337	0.00026
2020-09-17	111.5468	111.5463	0.00047
2020-09-18	112.3366	112.3360	0.00066
2020-09-19	112.0790	112.0791	8.9e-5
2020-09-20	111.3671	111.3678	0.00071
2020-09-21	111.1350	111.1356	0.00057
2020-09-22	111.9137	111.9138	2.0e-5
2020-09-23	113.5047	113.5044	0.00024
2020-09-24	115.1175	115.1171	0.00045
2020-09-25	115.9074	115.9067	0.00064
2020-09-26	115.6498	115.6499	0.00011
2020-09-27	114.9378	114.9385	0.00073
2020-09-28	114.7058	114.7063	0.00059
2020-09-29	115.4845	115.4845	4.1e-5
2020-09-30	117.0754	117.0752	0.00021
2020-10-01	118.6883	118.6879	0.00042
2020-10-02	119.4781	119.4775	0.00062
2020-10-03	119.2205	119.2206	0.00013
2020-10-04	118.5086	118.5093	0.00075
2020-10-05	118.2765	118.2771	0.00061
2020-10-06	119.0552	119.0553	6.2e-5
2020-10-07	120.6461	120.6459	0.00019
2020-10-08	122.2590	122.2586	0.00040
2020-10-09	123.0488	123.0483	0.00059
2020-10-10	122.7912	122.7914	0.00015
2020-10-11	122.0793	122.0801	0.00077
2020-10-12	121.8472	121.8479	0.00063
2020-10-13	122.6260	122.6260	8.4e-5
2020-10-14	124.2169	124.2167	0.00017
2020-10-15	125.8298	125.8294	0.00038
2020-10-16	126.6196	126.6190	0.00057
2020-10-17	126.3620	126.3622	0.00017
2020-10-18	125.6500	125.6508	0.00079
2020-10-19	125.4180	125.4186	0.00065
2020-10-20	126.1967	126.1968	0.00010
2020-10-21	127.7876	127.7875	0.00015
2020-10-22	129.4005	129.4001	0.00036
2020-10-23	130.1903	130.1898	0.00055
2020-10-24	129.9327	129.9329	0.00019
2020-10-25	129.2208	129.2216	0.00081
2020-10-26	128.9887	128.9894	0.00067
2020-10-27	129.7674	129.7676	0.00013
2020-10-28	131.3584	131.3582	0.00013
2020-10-29	132.9712	132.9709	0.00034
2020-10-30	133.7611	133.7605	0.00053
2020-10-31	133.5035	133.5037	0.00022
2020-11-01	132.7915	132.7923	0.00084
2020-11-02	132.5595	132.5601	0.00069
2020-11-03	133.3382	133.3383	0.00015
2020-11-04	134.9291	134.9290	0.00011
2020-11-05	136.5420	136.5417	0.00032
2020-11-06	137.3318	137.3313	0.00051
2020-11-07	137.0742	137.0744	0.00024
2020-11-08	136.3622	136.3631	0.00086
2020-11-09	136.1302	136.1309	0.00071
2020-11-10	136.9089	136.9091	0.00017
2020-11-11	138.4998	138.4997	8.8e-5
2020-11-12	140.1127	140.1124	0.00030
2020-11-13	140.9025	140.9021	0.00049
2020-11-14	140.6449	140.6452	0.00026
2020-11-15	139.9330	139.9339	0.00088

Timing

Median of all measured repetitions.

Step	Effect Prophet	Python Prophet	Effect Prophet is
Fit	5.09 s	4.30 s	1.2× slower
Predict	692 µs	3.50 ms	Not compared: different work
Save model to JSON	1.08 ms	33.8 ms	31× faster
Load model from JSON	4.42 ms	16.5 ms	3.7× faster
Simulate uncertainty	7.58 ms	153 ms	Not compared: different work
Load and simulate in a new process	198 ms	1.65 s	Not compared: different work

Some steps aren't compared, because the two libraries' public calls produce different outputs. Their times show what each call costs, not which library is faster.

• Predict: Python's predict on a posterior also computes every component's range over the draws, while Effect Prophet's predict returns point forecasts and leaves ranges to predictWithIntervals.
• Simulate uncertainty, Load and simulate in a new process: Python's predict(vectorized=False) also builds the point forecast and every component, while Effect Prophet's predictUncertainty returns only the ranges.

Settings

The case declaration both libraries ran.

{
  "kind": "map",
  "comparison": {
    "kind": "equivalent-posterior",
    "evidenceId": "posterior-nuts-v1"
  },
  "configuration": {
    "growth": "linear",
    "scaling": "absmax",
    "seasonalityMode": "additive",
    "holidaysMode": "additive",
    "changepoints": {
      "mode": "auto",
      "count": 25,
      "range": 0.8
    },
    "changepointPriorScale": 0.05,
    "seasonalities": [
      {
        "name": "weekly-custom",
        "periodDays": 7,
        "fourierOrder": 3,
        "priorScale": 10
      }
    ],
    "events": [],
    "regressors": []
  },
  "pythonOptimizer": {
    "algorithm": "Auto",
    "maxIterations": 10000,
    "newtonFallback": true,
    "sigFigs": 12
  },
  "uncertainty": {
    "seed": 19,
    "samples": 1000,
    "intervalWidth": 0.8,
    "output": "intervals",
    "algorithm": "scalar-continuous-time",
    "evidence": "docs/validation/uncertainty.md#evidence-classes",
    "samplerMemoryLimitBytes": 67108864
  },
  "mcmc": {
    "samples": 300,
    "seed": 19,
    "pointWidthFraction": 0.15
  }
}