predicted weekly demand skewed by one day

[andycarlsonuk]

I've recently set up a loudml model to forecast demand on a system that I work on.

I trained it with data for one calendar month - 1st to 30th August.

Forecasting for October I see that it has predicted the shape of the 7 day variation in our usual demand curve quite nicely... with one problem - the curve is 1 day behind where it should be. 11th Oct was a Sunday for example and should have the smallest daytime demand peak but the forecast has the smallest peak on Monday 12th.

My model json is below. The data store is influxdb and the input measurement is the hourly total HTTP elapsed time across all servers. The 'mean_' prefix for the feature name is misleading - sorry about that.

{
    "default_bucket": "barprod",
    "name": "http_conc_foo",
    "type": "donut",
    "features": [
        {
            "name": "mean_http_secs",
            "metric": "sum",
            "field": "http_total_secs",
            "measurement": "time_model",
            "match_all": [
                {"tag": "instance", "value": "foo.com"}
            ],
            "anomaly_type": "low_high",
            "default":0 
        }
    ],
    "run": {
        "detect-anomalies": true,
        "datasink": "loudmlout"
    },
    "bucket_interval": "1h",
    "interval": "1h",
    "offset": "2h",
    "max_evals": 10,
    "forecast": 24,
    "seasonality": {
        "daytime": true,
        "weekday": true
    },
    "span": 24
}

Thanks in advance.

[andycarlsonuk]

I should have said earlier that the forecasting was done in 1 day chunks.

I've tried modifying 'forecast' in the json to 170 (1 week), dropping the forecast output data from influxdb and then forecasting in bigger chunks. With a 2 day forecast chunk the forecast profile lags the day of week by 2 days.

At the moment the model seems to be replaying a crude 'echo' of the actual value from 1 or 2 days previously. When there are anomalies in the actuals these also produce an 'echo' in the forecast.

eval-model does not seem to show the same 'echo' behaviour and the 'forecast' profile fits with the expected pattern for the day of the week.