如何将时间序列趋势转换为可测量的预测变量

Question

I have a time series data which explains the number of frauds in the transaction over 1 year timeline along with the target variable of fraud or not.

X- axis is time-line and Y- axis is number of frauds detected.

Do we have any ML model/statistical technique that tries to identify the trend in these frauds and convert into a measurable predictor variable with value like 0 to 1, where values close to 1 are more prone to fraud and vic.,

The trends in the frauds over an year is non-linear, so if there is any mathematical transformation i can apply on the time-series so that it can provide me a measurable feature?

Any suggestions are much appreciated?

I thought of using normal slope techniques where negative slope wrt time-line are less fraud and positive slope for more fraud. It only captures linear trend, but need to capture non-linear trend.

Edit::

I forgot one important point. I will give one scenario to explain this point better.

For Financial banks, let's say I have 1000 banks and each bank has 12 months time period of how many frauds detected per month and corresponding target variable whether that bank has high chances of fraud or not.

Now, when I encounter a new bank with corresponding frauds in 12 months, what are the ways to find whether that bank is fraud or not using the 1000 banks fraud pattern?

Can we use any time-series approach? I assume, if it is for single bank, time-series handles it as I have multiple banks, I guess using non-linear regression techniques, assuming each month as one feature, training a model might help? As I can get a polynomial equation which I can use to predict the target?

Please share your thoughts as well

Answer 1

I'm going to assume your data includes risk variables (Customer data, loan data, etc.). I have used linear models, logistical models and conditional interference trees for this. The following is a vary high level view. You really need to understand underlying methods to get a good, functional model. I recommend the use of dummy variables with these...binary are best for easy interpretation.

A linear model or logistical model will result in an equation you can use to measure risk of each record (loan). This method requires kicking out outliers, Cook's distance, etc.

Linear:

step(lm(default flag ~ Variable 1 + Variable 2 + ...))

Logistical:

glm(default flag ~ Variable 1 + Variable 2 + ...)

The other is conditional interference trees. I would use partykit() package with the ctree() function. This will bucket defaults based on statistical significance of variables within the bucket.

plot(ctree(default flag ~ Variable 1 + Variable 2, data= "Your data", 
controls = "looks this up for your model")

Also, if you are worried about Time to default as well, then look into a survival analysis.

I have had success with all three. If all you have is a time period and default total, you can't really do much with that since you wouldn't have the underlying variables.