Mixed features, two chains

match ← All flat growth benchmarks

96 daily rows with min-max scaling, switchable multiplicative weekly seasonality, an event and a binary regressor. mcmc_samples=500 on 2 chains: 500 draws, two futures each.

Input

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

Output

Largest difference: 0.0062 (0.022% of the input's range).

Every forecast row from both libraries.

Date	Effect Prophet	Python Prophet	Difference
2020-04-06	0.2173	0.2112	0.0062
2020-04-07	7.1351	7.1322	0.0029
2020-04-08	36.2290	36.2324	0.0034
2020-04-09	62.8870	62.8875	0.00049
2020-04-10	35.0022	35.0022	3.5e-5
2020-04-11	50.4719	50.4768	0.0050
2020-04-12	19.5123	19.5127	0.00040
2020-04-13	0.2173	0.2112	0.0062
2020-04-14	8.3356	8.3341	0.0015
2020-04-15	35.0022	35.0022	3.5e-5
2020-04-16	62.8870	62.8875	0.00049
2020-04-17	69.7632	69.7646	0.0014
2020-04-18	50.4719	50.4768	0.0050
2020-04-19	19.5123	19.5127	0.00040
2020-04-20	36.2027	36.2042	0.0014
2020-04-21	7.1351	7.1322	0.0029
2020-04-22	35.0285	35.0305	0.0020
2020-04-23	62.8870	62.8875	0.00049
2020-04-24	69.7632	69.7646	0.0014
2020-04-25	35.0022	35.0022	3.5e-5
2020-04-26	20.7128	20.7147	0.0018
2020-04-27	0.2173	0.2112	0.0062
2020-04-28	7.1351	7.1322	0.0029
2020-04-29	35.0285	35.0305	0.0020

Timing

Median of all measured repetitions.

Step	Effect Prophet	Python Prophet	Effect Prophet is
Fit	636 ms	529 ms	1.2× slower
Predict	312 µs	4.90 ms	Not compared: different work
Save model to JSON	166 µs	12.5 ms	75× faster
Load model from JSON	1.51 ms	9.74 ms	6.5× faster
Simulate uncertainty	2.02 ms	130 ms	Not compared: different work
Load and simulate in a new process	185 ms	1.56 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": "flat",
    "scaling": "minmax",
    "seasonalityMode": "additive",
    "holidaysMode": "additive",
    "changepoints": {
      "mode": "auto",
      "count": 25,
      "range": 0.8
    },
    "changepointPriorScale": 0.05,
    "seasonalities": [
      {
        "name": "weekly-custom",
        "periodDays": 7,
        "fourierOrder": 2,
        "priorScale": 10,
        "conditionName": "active",
        "mode": "multiplicative"
      }
    ],
    "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": 500,
    "seed": 19,
    "chains": 2,
    "pointWidthFraction": 0.15
  }
}