0
$\begingroup$

I have 2 independent data sets (1. 300 rows and 2.3000 rows) with 6 months trades observations for 50 traders. In both datasets I have: trader id, stock title, buy/sell volume, date of trade, sector of stock

My goal is to detect possible outliers (suspicious trades) in this two datasets.

  1. What algorithm you would recommend for this 2 tasks and why?
  2. What can we use for 1 task when we have average only 5-6 trades per trader?
$\endgroup$
1
$\begingroup$

I'd go out on a limb to say this is a difficult problem without incorporating other data. Models just try to explain future action with past information. Traders act differently in different scenarios.

Without data on their analysis or financial data on the stocks they are analyzing, a model on the previous transaction history of traders would not be effective. Unless, if the "traders" you describe are more like etf fund managers that operate on a predefined set of rules.

| improve this answer | |
$\endgroup$
  • $\begingroup$ what other financial data would you like to include to solve those tasks? growth/fall of stock price? if you would get you data what algorithm would you use then $\endgroup$ – John Doe Jun 30 '19 at 14:14
  • $\begingroup$ It depends on the traders' strategy. If unknown, include all. Price action, security fundamental data, options pricing data on the underlying, etc. $\endgroup$ – Alexander Jun 30 '19 at 14:16
  • $\begingroup$ ok, thanks. And what model/algorith you would recommend for this? k-mean? something else? $\endgroup$ – John Doe Jun 30 '19 at 14:34
1
$\begingroup$

Regarding outliers,

  1. If your data is normally distributed, then using Z-Score value you can evaluate each feature and identify the outliers. You can also take the reference of Box-plot. All these techniques are based on Standard Deviation and help you in anomaly detection. More about Z-Score
  2. Scikit framework provides you interesting methods for outliers detection. If you can prepare your data in the form of inlier and outlier samples, then you can trained a model and evaluate any new sample for anomaly. Refer Scikit-Learn: Novelty and Outlier Detection

Hoping this will be helpful.

| improve this answer | |
$\endgroup$
0
$\begingroup$

what do you mean "outliers"? if you mean new behavior based on numbers then just go row by row, see if new observation is in some safe bounds based on historical data and update these bounds based on new observations (because trades volume may increase over time and you would like to take that into account), maybe some RNN would learn pattern of behavior, or you can look into some solutions of Kaggle contests where target was to identify suspicious behavior

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.