Business Cases of Regression Problems and how to approach them?

Looking at the business cases of Telecom, Retail, Finance, Health care, their Core Predictors, & the approaches to solve them!

Generated by Mid-Journey

Hello Everyone! The objective of this story is to examine the business problem first, then the problem statement, followed by the required data elements for predictions and regression functions and how they can help solve it.

We will take 2–3 business problem cases from different industries (Retail, Telecom, Finance, Health Care), examine their intricacies, and apply the appropriate regression solution.

Let’s look into the Business problems:

Retail/Manufacturing/Telecom: Predict Revenue, Sales, Costs, Demand.

A business problem could be improving the company's profit. All of these factors—revenue, sales, costs, and demand—help to achieve either more revenue or reduce costs to achieve or surpass the forecasted numbers and make the company profitable. Demand forecasts help to plan and maintain the optimal number of stock inventories (and to boost sales ) to meet market demand.

Finance: Predict Stock Risks/Volatility.

A business problem could be identifying stocks suitable for recommendation (e.g., Buy, Sell, Hold Calls), thereby enhancing customer engagement in a finance agency.

Health Care: Predict the health care of patients.

Business problems could be to predict the likelihood of recovery from the disease (Eg, COVID-19).

Let’s write the Problem Statements:

Retail/Manufacturing: Predict Daily demand/sales (eg: dairy products) so that products inventory can be maintained as only required.

Telecom: Predict Broadband Fibre, Digital TV services sales during coming christmas season or next month so that more network capacity can be added and investment planning as accordingly.

Finance: Predict stock volatility so that stock financial advisors can make the recommendations for their customers to invest on them.

Health Care: Predict recovery time of disease (eg: covid-19 recovery time) based upon treatment, inoculation history, patient medical history, age, personal & other factors.

Quick Look at Regression Types:

Linear Regression plots the line of best fit linearly based on the data point observations. This is a basic ML model, and we have many stories explaining it. I do not want to go in-depth about solving these solved models. We will look at our approach to the problem and choose the right one to fit it appropriately with our business cases.

The type of Regression depends upon the number of feature variables or predictors & their linear relationship function.

Finding out the relationships of dependent with Independent variable(s):

If one only independent variable for the prediction — Simple Linear Regression

y = mx + b + e whereas x is the independent variable used to predict; y is the dependent variable to predict; b is the y intercept; m is the slope of the explanatory variable; e is the residual error term

If more than independent variable — Multiple Linear Regression

y = m1x1 + m2x2 + m3x3 + … + mtxt + b + e whereas x1,x2..,xt are the independent variable used to predict; y is the dependent variable to predict; b is the y intercept; m1,m2,m3..mt are the slope of the explanatory variable; e is the residual error term

If the relationship equation having polynomials functions — Polynomial Regression

y = m1x1 + m2x1 ^2 + m3x1 ^3 + … + mtx1 ^t + b + e whereas x1 is the independent variable used to predict; y is the dependent variable to predict; b is the y intercept; m1,m2,m3..mt are the slope of the explanatory variable; e is the residual error term

Finding out the movement & relationship with time:

If the Predicted variable is same as independent, and variable moves by itself with time — Auto Regression.

This is the Time Series Model Approach. Many classical time series model intuiting from this basic concept with slightly different approaches (ARIMA, SARIMA, SARIMAX..) taking moving average or Autoregressive, or integrating them together, + adding seasonal component to it.

Choosing the right approach for our use cases:

Manufacturing: Predict Daily sales (eg: dairy products) so that products inventory can be maintained as only required.

List of characteristics of a dairy farm that can impact the daily production quantity or the sale quantity:

Generated by Mid-Journey

1. The Land Area of the farm
2. Farm Size — Large/Medium/Small
3. Number of cows
4. Product Type: Milk/Yoghurt/Cheese
5. Shelf Life: Number of days the product remains usable
6. Storage Condition: Frozen/Refrigerated

Solution: Multiple Linear Regression Approach to Predict Daily Sales Quantity — Just fitting these features in the Multiple Linear Regression function above (x1,x2,x3..)

Retail: Predict Daily demand(eg: dairy products) so that products inventory can be maintained as only required.

List of characteristics of a dairy demand for a retail store that can impact the daily production quantity or the sale quantity:

Generated by Mid-Journey

1. Closeby Retail Stores availability.
2. Age group bin count of people living around the store.
3. Customers' ethnicity — A PII variable demands using them as per governance policies.
4. Product Type
5. Shelf Life: Number of days the product remains usable
6. Storage Condition: Frozen/Refrigerated
7. Number of Customers visiting the store/online purchases, ..
8. Customer Demographics

Solution: Multiple Linear Regression Approach to Predict Daily Sales Quantity — Just fitting these features in the Multiple Linear Regression function above (x1,x2,x3..)

Get M. Suresh Kumar’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

This problem can also be attacked with time series (Auto Regression and a combination of other approaches are also possible). Experiment results prove the best fit.

Telecom: Predict Broadband Fibre, Digital TV services sales during coming christmas season or next month so that more network capacity can be added and investment planning as accordingly.

List of characteristics of a new BB service sale of a Telecom company that helps the company more profitable, and make the company informed for any investments, and increase the network infrastructure, helping to improve customer engagement & satisfaction:

1. Competitor products price.
2. Competitor product features.
3. Customer needs & wants (e.g., Some customers are looking for medical and security alarms — BB service should be able to support the specific needs).
4. Customer 360 Data — Age, Demographics, Income & affordability, Customer Affinity & traits.

Solution: Multiple Linear Regression Approach to predict BB Sales for the future — Just fitting these features in the Multiple Linear Regression function above (x1,x2,x3..).

Finance: Predict stock volatility that pave a way to make the recommendations for their customers to invest on them.

Beta helps to calculate the expected return of investments in a stock by applying the CAPM (capital asset pricing model) calculation (calculated values are in the CAGR % per year).

As we all know Beta — is one of the most sought after metric to pick a stock to invest — If Beta is higher than 1, the volatility of the stock is higher than the market & also the risk, If Beta is lower than 1, then the stock price volatility lower than the market and are stable and not change drastically with the market movements. If Beta is equal to 1, then the stock price volatility is as same as market volatility.

Beta Calculation Formula using covariance:

Beta=Covariance of the stock with the Market returns / Variance of the market returns 
Beta Coefficient = Covariance (Rn , Rm) /Variance (Rm)

whereas Rn - Return on the Individual Stock
Rm - Return on the Overall Market

List of characteristics to predict Beta value:

1. Competitor stocks price.
2. International market fluctuations.
3. Gold Prices & International Currency value fluctuations
4. Company P/E Price Earnings Ratio
5. Company Market capitalization
6. Net Profit, Net Revenue
7. Debt
8. Customer Base & Market of the Company
9. Competitor performance
10. Government Policies
11. Many other External factors market are impacted

Solution: Since many factors are involved, we can solve this stochastically. The multiple linear regression approach can be one approach to predicting a stock's Beta—just fit these features in the multiple linear regression function above (x1,x2,x3...).

Also, this can be considered as Non-Linear, and can use more advanced algorithms to solve.

CAPM Calculation (Expected Rate of Investment Return):

Expected Rate of Investment Return = Risk Free Return + Beta (Market Return)
Market Return = Expected Rate of Return based on Market Premium- Risk Free Return
Note: These are CAGR % for an year
Beta,& the Expected Rate of Return based on Market premium are the primary variable components.
Risk Free Return is more or less flat.

Risk Free Return - can be the fixed deposit returns
Market Return - is the return for the risk taken by investing in a stock - that's why it is a premium return

Health Care: Predict recovery time of disease (eg: covid-19 recovery time) based upon treatment, inoculation history, patient medical history, age, personal & other factors.

List of characteristics of a patient for predicting the recovery time

Generated by Mid-Journey

1. Domicile of Patient.
2. Age.
3. Gender
4. Comorbidities
5. Diabetes
6. Hypertension
7. Coronary Artery Disease
8. Inoculation History

Solution: Multiple Linear Regression Approach to predict disease recovery time — Just fitting these features in the Multiple Linear Regression function above (x1,x2,x3..).

Again, all of these MLR cases can also be converted to a classification problem instead of a Regression problem by converting the target variable to a discrete one instead of a continuous variable. They can also be solved using advanced Non-Linear approaches, such as XGBoost, GBTs, Random Forest, ANN, etc., which mostly yield better results (as a classification or Regression problem).

Also, this can be the problem candidate for AutoML (Use Google, Azure, or AWS cloud if you want to try AutoML quickly) to select the best model that gives the best results.

We have reached the end of the story. To optimize these models, especially the Loss and error functions, we need more advanced concepts. We will cover them in the next stories. Stay tuned!

If you like this story and want to subscribe for more, click below https://medium.com/@mskmiba/subscribe

Follow for more such content around data & analytics, Social Interest, and Well-being.

Thanks for reading!

Please follow, clap, and support me so that I can write more. Cheers!

In Plain English 🚀

Thank you for being a part of the In Plain English community! Before you go:

• Be sure to clap and follow the writer ️👏️️
• Follow us: X | LinkedIn | YouTube | Discord | Newsletter | Podcast
• Create a free AI-powered blog on Differ.
• More content at PlainEnglish.io