tfp.sts future forecasting with external (holiday) regressor

[abc-8]

Referring to electricity demand in the example notebook:

https://github.com/tensorflow/probability/blob/master/tensorflow_probability/examples/jupyter_notebooks/Structural_Time_Series_Modeling_Case_Studies_Atmospheric_CO2_and_Electricity_Demand.ipynb

given the future temperature (e.g. 2014-02-27 00:00:00 -> 2014-02-28 00:00:00), how can I
call tfp.sts.forecast method?

[davmre]

Great question. It's a little confusing to talk about because of a distinction we didn't make very clear in the post: of course a model that depends on external covariates can't properly produce 'forecasts' of the future (because it requires we know the values of the covariates!) -- but it can be used for 'nowcasting' (estimating a hard-to-observe or infrequently-observed quantity given an easily-observed quantity, e.g., flu occurrence given web search trends), or of course you can go deeper and plug in forecasts of the covariates, though properly handling the uncertainty in these forecasts can gets tricky. Concretely: `tfp.sts.forecast` extends each model component over additional steps, including the LinearRegression component, so it needs a model that includes the 'future' covariates as part of the LinearRegression design matrix. That is, the design matrix should concatenate the temperature for both the training and 'forecast' periods: ``` # training_temperature.shape == [num_train_steps, 1] # forecast_temperature.shape == [num_forecast_steps, 1] temperature_effect = tfp.sts.LinearRegression( design_matrix = tf.concat([training_temperature, forecast_temperature], axis=0)) ``` This is actually what the example notebook already does, implicitly: it just defines 'temperature' that covers both the training and forecast time periods. Hope this helps! Dave

…

On Fri, Mar 29, 2019 at 7:42 AM abc-8 ***@***.***> wrote: Referring to electricity demand in the example notebook: https://github.com/tensorflow/probability/blob/master/tensorflow_probability/examples/jupyter_notebooks/Structural_Time_Series_Modeling_Case_Studies_Atmospheric_CO2_and_Electricity_Demand.ipynb given the future temperature (e.g. 2014-02-27 00:00:00 -> 2014-02-28 00:00:00), how can I call tfp.sts.forecast method? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#343>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA8iitZW8g-7bhipBT65_bqxEpvoay12ks5vbiZkgaJpZM4cSi4o> .

[abc-8]

@davmre Thanks,

what I meant is how you continue forecasting the future after (production), say you have a weather service that feeds your production model with tomorrow's hourly weather, so you always have to append tomorrow weather to forecast tomorrow's demand.

I found a solution for this by hacking. that is, mutating LinearRegression's design_matrix before calling tfp.sts.forecast(), and its working like a charm:

# in this example, consider the dataset contains demand data till today
tomorrow_weather = service.weather_future_forecast('tomorrow') # next 24 hours
weather_matrix = tf.concat([ temperature_effect._design_matrix.to_dense(), tomorrow_weather], axis=0)
temperature_effect._design_matrix = tf.linalg.LinearOperatorFullMatrix(
            weather_matrix,
            name='design_matrix_linop')
# continue by calling tfp.sts.forecast with `num_steps_forecast=24`

and that lead us to another question:

is this the right way to forecast ? or do I need to retrain the model everytime I change LinearRegression design_matrix ? I don't have extensive knowledge about variational inference that's why I'm wondering if this is OK.

Thanks,

[davmre]

Mutating private members is probably not an engineering best practice, but mathematically what you're doing is fine. Fitting the model (whether with VI or HMC) gives you a set of posterior samples for the parameters of the model given your training inputs; for LinearRegression, those parameters are the regression weights. As long as your new design matrix has the same set of features, there's no reason to retrain new weights unless you want to incorporate new training data. From a coding standpoint, you could avoid mutating private state by just creating a new STS model object (including the LinearRegression and whatever other components you're using) with the new design matrix. STS models are stateless, so once you've gotten posterior samples of the parameters, there's nothing special about the original model object; a new model with the same structure (thus same set of parameters) but a different design matrix will work just as well. Dave

…

On Mon, Apr 8, 2019 at 7:29 AM abc-8 ***@***.***> wrote: @davmre <https://github.com/davmre> Thanks, what I meant is how you continue forecasting the future after (production), say you have a weather service that feeds your production model with tomorrow's hourly weather, so you always have to append tomorrow weather to forecast tomorrow's demand. I found a solution for this by hacking. that is, mutating LinearRegression's design_matrix before calling tfp.sts.forecast(), and its working like a charm: # in this example, consider the dataset contains demand data till today tomorrow_weather = service.weather_future_forecast('tomorrow') # next 24 hours weather_matrix = tf.concat([ temperature_effect._design_matrix.to_dense(), tomorrow_weather], axis=0) temperature_effect._design_matrix = tf.linalg.LinearOperatorFullMatrix( weather_matrix, name='design_matrix_linop')# continue by calling tfp.sts.forecast with `num_steps_forecast=24` and that lead us to another question: *is this the right way to forecast ? or do I need to retrain the model everytime I change LinearRegression design_matrix ?* I don't have extensive knowledge about variational inference that's why I'm wondering if this is OK. Thanks, — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#343 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA8iiiaM7ZuHlhN6aYaLS-7tTXFA43X5ks5ve1JXgaJpZM4cSi4o> .

[axch]

@abc-8, assuming your question has been answered, I'm going to go ahead and close this issue. Please reopen (or file a new one) as needed.