Mixed features, 600 draws

match ← All linear growth benchmarks

96 daily rows with switchable multiplicative weekly seasonality, an event, a binary regressor and two automatic changepoints. mcmc_samples=300: 4 chains of 150 draws, and 80% ranges from 1,200 simulated futures, two per draw, on 24 rows.

Input

benchmark/inputs/v1/linear-mixed-components.json: 96 history rows and 24 forecast rows.

Output

Largest difference: 0.0069 (0.019% of the input's range).

Every forecast row from both libraries.

Date	Effect Prophet	Python Prophet	Difference
2020-04-06	26.0165	26.0168	0.00028
2020-04-07	29.3869	29.3873	0.00048
2020-04-08	44.1649	44.1664	0.0015
2020-04-09	56.3159	56.3195	0.0036
2020-04-10	42.9831	42.9856	0.0025
2020-04-11	50.5385	50.5423	0.0038
2020-04-12	35.6511	35.6536	0.0025
2020-04-13	26.3563	26.3573	0.00093
2020-04-14	31.1080	31.1084	0.00039
2020-04-15	43.3811	43.3844	0.0033
2020-04-16	57.0473	57.0523	0.0050
2020-04-17	60.5083	60.5138	0.0054
2020-04-18	51.1925	51.1975	0.0050
2020-04-19	36.1115	36.1150	0.0034
2020-04-20	45.1319	45.1352	0.0033
2020-04-21	30.1530	30.1549	0.0019
2020-04-22	43.9564	43.9609	0.0045
2020-04-23	57.7787	57.7851	0.0065
2020-04-24	61.2827	61.2896	0.0069
2020-04-25	44.1771	44.1819	0.0048
2020-04-26	37.9396	37.9431	0.0035
2020-04-27	27.0359	27.0381	0.0022
2020-04-28	30.5361	30.5388	0.0027
2020-04-29	44.5139	44.5194	0.0055

Timing

Median of all measured repetitions.

Step	Effect Prophet	Python Prophet	Effect Prophet is
Fit	139 ms	186 ms	1.3× faster
Predict	288 µs	5.07 ms	Not compared: different work
Save model to JSON	273 µs	13.0 ms	48× faster
Load model from JSON	2.62 ms	9.23 ms	3.5× faster
Simulate uncertainty	2.95 ms	164 ms	Not compared: different work
Load and simulate in a new process	187 ms	1.64 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": "multiplicative",
    "holidaysMode": "additive",
    "changepoints": {
      "mode": "auto",
      "count": 2,
      "range": 0.8
    },
    "changepointPriorScale": 0.2,
    "seasonalities": [
      {
        "name": "weekly-custom",
        "periodDays": 7,
        "fourierOrder": 2,
        "priorScale": 10,
        "conditionName": "active"
      }
    ],
    "events": [
      {
        "name": "campaign",
        "date": "2020-01-22",
        "lowerWindowDays": 0,
        "upperWindowDays": 0,
        "priorScale": 10
      }
    ],
    "regressors": [
      {
        "name": "promotion",
        "priorScale": 10,
        "standardization": "never"
      }
    ]
  },
  "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
  }
}