Methods and systems for generating predictions based on time series data using an ensemble modeling
The methods and systems provide an ensemble approach that combines multiple single-model-solutions to produce optimal forward-looking forecasts. Moreover, the methods and systems provide an architecture for this ensemble approach that ensures that the limitations for individual ensemble model components are compensated by other ensemble model components as inputs and outputs from ensemble model components are fed from one ensemble model component to another in a specific order to generate a final output upon which a conservative prediction is based.
1. A system for using an ensemble machine learning model to generate predictions based on time series data, the system comprising:
cloud-based storage circuitry configured to store the ensemble machine learning model, wherein the ensemble machine learning model is trained on a plurality of samples of time series data, wherein the plurality of samples of time series data is generated based on cross-validation resampling;
control circuitry configured to:
receive a first feature input, wherein the first feature input represents a first set of times series data;
input the first feature input into a first ensemble model component and a second ensemble model component, wherein the first ensemble model component is a mean-based regression model, and wherein the second ensemble model component is a quantile regression model;
receive a first output from the first ensemble model component and a second output from the second ensemble model component, wherein the first output is based on a point estimate based on the first feature input, and wherein the second output is based on a lower tail distribution of a quantile regression;
generate a second feature input based on the first output and the second output, wherein generating the second feature input based on the first output and the second output further comprises:
generating a plurality of respective outputs from a plurality of ensemble model components, wherein respective inputs for the plurality of ensemble model components comprise the first output or the second output; and
determining the second feature input based on the respective outputs;
input the second feature input into a final ensemble model component to determine a final output, wherein the final ensemble model component comprises the plurality of ensemble model components used to generate predictions subject to a constraint; and
input/output circuitry configured to generate for display, on a user interface, a prediction based on the final output.
2. A method for using an ensemble machine learning model to generate predictions based on time series data, the method comprising:
receiving a first feature input, wherein the first feature input represents a first set of times series data;
inputting the first feature input into a first ensemble model component and a second ensemble model component, wherein the first ensemble model component is a mean-based regression model, and wherein the second ensemble model component is a quantile regression model;
receiving a first output from the first ensemble model component and a second output from the second ensemble model component;
generating a second feature input based on the first output and the second output;
inputting the second feature input into a final ensemble model component to determine a final output, wherein the final ensemble model component comprises a plurality of ensemble model components used to generate predictions subject to a constraint, and wherein generating the predictions subject to the constraint comprises imposing an upper bound on over-estimation frequency and an upper bound on a maximum overestimation value; and
generating for display, on a user interface, a prediction based on the final output.
3. The method of claim 2 , wherein the first output is based on a point estimate based on the first feature input, and wherein the second output is based on a lower tail distribution of a quantile regression.
4. The method of claim 2 , wherein generating the predictions subject to the constraint comprises minimizing a loss function while imposing a penalty for overestimation.
5. The method of claim 2 , further comprising:
determining respective weights for each of the plurality of ensemble model components; and
applying the respective weights to the final ensemble model component as parameters.
6. The method of claim 2 , wherein generating the second feature input based on the first output and the second output further comprises:
generating, based on the first output, a third feature input and a fourth feature input;
inputting the third feature input into a third ensemble model component to determine a third output, wherein the third ensemble model component determines a downward shift for a point estimate, and wherein the downward shift is subject to a floor; and
inputting the fourth feature input into a fourth ensemble model component to determine a fourth output, wherein the fourth ensemble model component determines a standard deviation for past values of the point estimate.
7. The method of claim 6 , wherein generating the second feature input based on the first output and the second output further comprises:
generating, based on the second output, a fifth feature input; and
inputting the fifth feature input into a fifth ensemble model component to determine a fifth output, wherein the fifth ensemble model component performs a quantile regression.
8. The method of claim 2 , wherein generating the second feature input based on the first output and the second output further comprises:
generating a plurality of respective outputs from the plurality of ensemble model components, wherein respective inputs for the plurality of ensemble model components comprise the first output or the second output; and
determining the second feature input based on the respective outputs.
9. The method of claim 2 , wherein the constraint is measured by minimizing a loss function based on mean absolute error or root mean squared error.
10. The method of claim 2 , further comprising:
receiving a second feature input, wherein the second feature input represents a second set of time series data;
generating a plurality of samples based on cross-validation resampling of the second set of time series data; and
training the ensemble machine learning model based on the plurality of samples.
11. A non-transitory, computer-readable medium for generating hierarchical data structures based on crowdsourced data featuring non-homogenous metadata, comprising instructions that, when executed by one or more processors, cause operations comprising:
receiving a first feature input, wherein the first feature input represents a first set of times series data;
inputting the first feature input into a first ensemble model component and a second ensemble model component, wherein the first ensemble model component is a mean-based regression model, and wherein the second ensemble model component is a quantile regression model;
receiving a first output from the first ensemble model component and a second output from the second ensemble model component;
generating a second feature input based on the first output and the second output;
inputting the second feature input into a final ensemble model component to determine a final output, wherein the final ensemble model component comprises a plurality of ensemble model components used to generate predictions subject to a constraint, and wherein generating the predictions subject to the constraint comprises imposing an upper bound on over-estimation frequency and an upper bound on a maximum overestimation value; and
generating for display, on a user interface, a prediction based on the final output.
12. The non-transitory, computer-readable medium of claim 11 , wherein the first output is based on a point estimate based on the first feature input, and wherein the second output is based on a lower tail distribution of a quantile regression.
13. The non-transitory, computer-readable medium of claim 11 , wherein generating the predictions subject to the constraint comprises minimizing a loss function while imposing a penalty for overestimation.
14. The non-transitory, computer-readable medium of claim 11 , further comprising:
determining respective weights for each of the plurality of ensemble model components; and
applying the respective weights to the final ensemble model component as parameters.
15. The non-transitory, computer-readable medium of claim 11 , wherein generating the second feature input based on the first output and the second output further comprises:
generating, based on the first output, a third feature input and a fourth feature input;
inputting the third feature input into a third ensemble model component to determine a third output, wherein the third ensemble model component determines a downward shift for a point estimate, and wherein the downward shift is subject to a floor; and
inputting the fourth feature input into a fourth ensemble model component to determine a fourth output, wherein the fourth ensemble model component determines a standard deviation for past values of the point estimate.
16. The non-transitory, computer-readable medium of claim 15 , wherein generating the second feature input based on the first output and the second output further comprises:
generating, based on the second output, a fifth feature input; and
inputting the fifth feature input into a fifth ensemble model component to determine a fifth output, wherein the fifth ensemble model component performs a quantile regression.
17. The non-transitory, computer-readable medium of claim 11 , wherein generating the second feature input based on the first output and the second output further comprises:
generating a plurality of respective outputs from the plurality of ensemble model components, wherein respective inputs for the plurality of ensemble model components comprise the first output or the second output; and
determining the second feature input based on the respective outputs.
18. The non-transitory, computer-readable medium of claim 11 , wherein the constraint is measured by minimizing a loss function based on mean absolute error or root mean squared error.