Machine learning-based temporal startup predictive system

Patent metadata

Field	Value
Publication number	US-12315010-B2
Application number	US-202217828386-A
Country	US
Kind code	B2
Filing date	May 31, 2022
Priority date	May 31, 2022
Publication date	May 27, 2025
Grant date	May 27, 2025

Abstract

Systems and methods are directed to predicting temporal startup measurements using a machine-trained model. The system determines a training dataset of features associated with different funding, exit, and closure events and corresponding times of the funding, exit, and closure events from historical financial data. A temporal prediction model is trained using the training dataset. The temporal prediction model can comprise a recurrent neural network (e.g., gated recurrent unit). During runtime, the system accesses new data associated with potential future investment opportunities with startups and determines (e.g., compute) company features based, in part, on the new data. The system applies the company features to the temporal prediction model to simultaneously predict a next event and a time of the next event for each startup. A user interface can then be presented that shows the predicted next event and the predicted time of the predicted next event for each startup.

First claim

What is claimed is:

1. A method comprising:
determining a training dataset of features associated with different funding, exit, and closure events and corresponding times of the funding, exit, and closure events from historical financial data;
training a temporal prediction model using the training dataset, the temporal prediction model comprising a recurrent neural network that is a gated recurrent unit, a marker prediction module to predict a next event, and a time prediction module to predict a time of the next event;
during runtime, accessing new data associated with potential future investment opportunities with startups;
determining company features based, in part, on the new data;
simultaneously predicting the next event and the time of the next event for each startup by applying the company features to the temporal prediction model, the simultaneously predicting comprising:
mapping event labels and timing information included in the company features to a hidden representation by the recurrent neural network; and
using the hidden representation as a feature by the marker prediction module to predict the next event and by the time prediction module to predict the time of the next event for each startup; and
causing presentation of a user interface that presents the predicted next event and the predicted time of the predicted next event for each startup.

2. The method of claim 1 , wherein:
the new data comprises a batch of data in a comma-separated values (CSV) file format; and
the determining the company features comprises feeding the batch of data through a featurizer that determines the company features.

3. The method of claim 1 , wherein determining the company features comprises determining funding event data from the new data for each startup, the funding event data including a current funding event, any previous funding events, and time of each funding event.

4. The method of claim 1 , wherein determining the company features comprises determining one or more momentum metrics for each startup, the one or more momentum metrics indicating momentum for each startup based on interest and growth indicators.

5. The method of claim 1 , wherein determining the company features comprises determining, by a featurizer, one or more derivative metrics for each startup by performing a derivative calculation using the new data.

6. The method of claim 5 , wherein the one or more derivative metrics includes a funding velocity.

7. The method of claim 1 , further comprising:
generating a marker embedding from one or more markers indicating one or more events to be predicted by the temporal prediction model; and
concatenating the marker embedding with the company features prior to applying the company features to the temporal prediction model.

8. The method of claim 1 , wherein the causing presentation of the user interface comprises causing presentation of a dashboard user interface that comprises a plurality of rows, each row indicating a startup name, a current funding event, the predicted next event, and the time of the next event.

9. The method of claim 1 , further comprising:
verifying accuracy of the temporal prediction model using the historical financial data.

10. The method of claim 1 , further comprising:
receiving new historical financial data; and
retraining the temporal prediction model using the new historical financial data.

11. The method of claim 1 , wherein the new data includes, for each startup, one or more of an amount of company queries, an amount of product queries, a number of job postings, message counts associated with interest in the job postings, or usage data on a cloud computing platform.

12. The method of claim 1 , wherein determining the company features comprises computing manually defined company features.

13. A system comprising:
one or more hardware processors; and
a memory storing instructions that, when executed by the one or more hardware processors, cause the one or more hardware processors to perform operations comprising:
determining a training dataset of features associated with different funding, exit, and closure events and corresponding times of the funding, exit, and closure events from historical financial data;
training a temporal prediction model using the training dataset, the temporal prediction model comprising a recurrent neural network that is a gated recurrent unit, a marker prediction module to predict a next event, and a time prediction module to predict a time of the next event;
during runtime, accessing new data associated with potential future investment opportunities with startups;
determining company features based, in part, on the new data;
simultaneously predicting the next event and the time of the next event for each startup by applying the company features to the temporal prediction model, the simultaneously predicting comprising:
mapping event labels and timing information included in the company features to a hidden representation by the recurrent neural network; and
using the hidden representation as a feature by the marker prediction module to predict the next event and by the time prediction module to predict the time of the next event for each startup; and
causing presentation of a user interface that presents the predicted next event and the predicted time of the predicted next event for each startup.

14. The system of claim 13 , wherein:
the new data comprises a batch of data in a comma-separated values (CSV) file format; and
the determining the company features comprises feeding the batch of data through a featurizer that determines the company features.

15. The system of claim 13 , wherein determining the company features comprises determining funding event data from the new data for each startup, the funding event data including a current funding event, any previous funding events, and time of each funding event.

16. The system of claim 13 , wherein determining the company features comprises:
determining one or more momentum metrics for each startup, the one or more momentum metrics indicating momentum for each startup based on interest and growth indicators; or
determining, by a featurizer, one or more derivative metrics for each startup by performing a derivative calculation using the new data.

17. The system of claim 13 , wherein the operations further comprise:
generating a marker embedding from one or more markers indicating one or more events to be predicted by the temporal prediction model; and
concatenating the marker embedding with the company features prior to applying the company features to the temporal prediction model.

18. A non-transitory machine-storage medium comprising instructions which, when executed by one or more hardware processors of a machine, cause the machine to perform operations comprising:
determining a training dataset of features associated with different funding, exit, and closure events and corresponding times of the funding, exit, and closure events from historical financial data;
training a temporal prediction model using the training dataset, the temporal prediction model comprising a recurrent neural network that is a gated recurrent unit, a marker prediction module to predict a next event, and a time prediction module to predict a time of the next event;
during runtime, accessing new data associated with potential future investment opportunities with startups;
determining company features based, in part, on the new data;
simultaneously predicting the next event and the time of the next event for eac

Assignees

• Microsoft Technology Licensing Llc

Inventors

• Bijral Avleen Singh
• Feldman Lee Eric
• Kumar Samir

Classifications

• G06N5/022

  Knowledge engineering; Knowledge acquisition · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06Q40/06 Primary

  Asset management; Financial planning or analysis · CPC title