Open for Business
Analyzing, modelling, and predicting the volume of new business applications in US states for July 2026.
Many of us have had a late-night brainstorming session that has led to a ‘genius’ business idea. But not everyone dares to convert these ideas into registered businesses. In this article, we aim to analyze data concerning those who take that leap and make business applications.
How much does location affect the volume of business applications for a specific month of the year? What about the various types of business applications; which ones recur the most? If given a specific month, how closely does it correlate with the preceding months of the year? Given data up to June 2026, can we predict the numbers for July 2026?
Using the latest Business Formation Statistics provided by the US Census Bureau, I have tried to answer these questions. The dataset spans from July 2004 to June 2026 and has data from 52 US states and territories concerning various business application types.
I. Analysis
There are 4 business application types taken into account: normal business applications that consist of a basic Employer Identification Number (EIN) application; these are abbreviated as BA. Other types are Business Applications from Corporation (CBA), High-propensity business applications (HBA), and, lastly, Business applications with planned wages (WBA).
Adding up the number of business applications in July of the preceding year (2025) reveals that Florida, California, Texas, New York, and Georgia are the 5 states with the largest application volumes across all types. Furthermore, the top 10 states make up 57.5% of the total volume of business applications.
If we consider the different business types for July across all years, we find that the BA type is by far the most frequent, with HBA in second. We can also observe a significant rise in applications in 2020 that quickly dipped in 2021.
To get a better sense of the volume of applications in July relative to the months preceding it (Jan-Jun), the heatmap below shows the correlation across all months from January to July for the years 2022–2025.
We can see that the correlation is very strong across the board, with the smallest coefficient being 0.9665. July has been especially strongly correlated with April, February, and January for the past 4 years.
Given this, we can also observe the volume of applications for all the months of 2026 that have already been published. The resulting graph shows that April has the largest number of applications so far in 2026.
II. Modelling & Predicting
To train the model, I used data from 2005 to 2021. The remaining data (2022–2025) were used to test the robustness of the model.
Get Humura Bliss’s stories in your inbox
Join Medium for free to get updates from this writer.
According to the Shapley values, January was the feature with the most importance in the model’s training. The other features are all ordered in the following graph according to importance:
Further evaluation of the model showed that the training RMSE is 866.19 units, while the testing RMSE is 1254.404 units, so overfitting was minimal.
The R2 score of the model on the testing set was 0.9634, meaning that the independent variables can explain 96.34% of the variance in the target variable (July).
Lastly, the model has an error with an average magnitude of 364.7 units; this represents a 10.4% average absolute error between predicted and actual values.
This model is, overall, a strong model, but it also has its weaknesses. One of the weaknesses is the model not being able to extrapolate and predicting the same number for larger volumes. This is due to the rarity of large volumes (over 40,000 applications in July) in the dataset, and the tuning of hyperparameters limited this problem but did not eliminate it.
The bar graph below is for the year 2026, with the model forecasting similar trends for July as other months: disproportionately BA-type applications, with HBA-type coming in second.
Furthermore, the model predicts that California will surpass Florida in the percentage of business applications, that Colorado will take the number 9 spot from Pennsylvania — which will drop from the top 10, and that the top 10 states will decrease from 57.5% to 54.2% of the total volume of applications.
Conclusion
In this article, I presented to you the way the volume of business applications in a month can be impacted by location, application type, and preceding months, among other features.
We saw that a majority of the applications are concentrated in 10 of the 52 states and territories represented and that BA application types are, by far, the most prevalent type of application. We also noticed a strong correlation between the preceding months and the target month.
We then looked at a strong model that predicted the values of July and used the model to predict the values for July 2026, which has not yet been released as of the date of publication of this article.
To see more of the code that gave the results discussed in this article, visit my GitHub here.
