Statistics of Polish law courts
about project
In July 2022 I intensely trained for Google TensorFlow Developer Certificate exam (practical 5 hour test in creating AI models of deep learning). This required of me gaining much experience in data processing beforehand. I proceeded on the assumption that it would be easier for me to work with the records I was best familiar with. As the Polish attorney at law with years of trial experience, I chose the data of Polish judicial system. This project is a "by-product" of my training.
What is this project about? It is an experiment for educational purposes. It presents selected data of the Polish judiciary. Based on this data, an artificial intelligence model can be trained directly in a browser. Then you can see what predictions such AI makes. Predictions are made about the length of court proceedings and the distribution of appellate court rulings. I hope that interacting with the project will be entertaining for you, as well as expand your knowledge of justice statistics and machine learning.
Why models are trained directly in the browser? For educational purposes. In result you are delivered live experience of adjusting the algorithm to the selected data. This is more exciting than making predictions with already trained model. If you would like to discuss details or give me any suggestions about the project I invite you to communicate with me through social media.
Tadeusz Mięsowicz
predict the average duration of legal proceedings
Based on datasets from the Ministry of Justice, the average duration of court proceedings in 2022 is predicted. To begin with, you can read explanations related to collected data. Training and prediction can be performed both for a specific category of proceedings and for all types of court proceedings. The categories of cases presented below were selected by me (there are others). These are typical civil and business proceedings involving plaintiffs and defendants.
predicting data based on one type of historical data
In this case we make predictions on the basis of a specific kind of historical data, the hitherto average durations of legal proceedings. This method, although experimentally exciting, may lead to some distortions of the picture. Taken alone, the general tendency of durations in question is not fully diagnostic about the future. Instead, one can produce models that take into account a wider number of variables in predicting the duration of legal proceedings. Training such models in a browser is completely inefficient.
types of data under analysis
Data categories are determined by the types of case repertories.
SO rep. C
The average duration (in months) of proceedings in civil cases before district courts.
SR rep. C
The average duration (in months) of proceedings in civil cases before regional courts.
SO rep. GC
The average duration (in months) of proceedings in business cases before district courts.
SR rep. GC
The average duration (in months) of proceedings in business cases before regional courts.
In toto
The average duration (in months) of all kinds of legal proceedings before common courts.
Available historical data (2011-2021)
The diagram below shows the data used to train AI.
first quarter of 2022 – historical data
Partial historical data available from 2022.
SO rep. C
10.75 months
SR rep. C
16.82 months
SO rep. GC
20.06 months
SR rep. GC
20.53 months
In toto
5.52 months
choose the category of data to be used for training & prediction
Press the suitable button to make your choice.
train AI and predict
The training lasts 500 epochs, i.e. the cycle of corrections of the model’s accuracy. The entire training should not take more than 5 minutes. The resulting prediction of duration for 2022 will be given automatically afterword.
AI training on your mobile
Because the computing workload I recommend not to train the model with phones, especially the older ones.
Number of epochs500
prediction result
After AI is trained you will see the result of its work.
The prediction for the entire 2022
No prediction yet
First quarter of 2022 – historical data
10.75 months
appeals
The analysis of data concerning handling the appeals by common courts permits to discover a correlation between the number of appeals brought before courts and the number of rulings of a particular kind. Let's see how machine learning will deal with these correlations. On the basis of data sets from the Ministry of Justice related to appeals, the annual number of specific substantive rulings of the courts of second instance are predicted. On top of this, the annual number of pending cases left for future years are predicted. In other words, on the basis of the annual number of appeals brought, AI predicts the annual number of rulings belonging to each category of decision (such as dismissal of appeals).
Regression
From the point of view of machine learning in the analysed case, we are dealing with regression. What is regression? To make things simple, the idea of regression is forecasting the continuous values of some variables on the basis of other continuous variables. That is, determining the value of a variable or variables using our knowledge of other variables. In our example, AI will infer the annual distribution of court’s rulings by their kinds from the annual number of appeals brought.
types of data under analysis
Data categories are determined by the types of case repertories.
Ca
The data concerning the yearly number of appeals and kinds of appellate rulings in civil cases before district courts.
Ga
The data concerning the yearly number of appeals and kinds of appellate rulings in business cases before district courts.
ACa
The data concerning the yearly number of appeals and kinds of appellate rulings in civil cases before appeals courts.
types of rulings - variables
Check out the data.
Appeals brought
Cases brought before appeal courts
Denied
Annual number of judgments denying appeals.
Changed
Annual number of judgments allowing appeals.
Sent back to I
Annual number of judgments resulting sending back case to the first instance.
Returned
Annual number of appeals not decided on the merits due to failure to complete formal requirements.
Rejected
Annual number of appeals not decided on the merits due to lack of formal case handling capabilities.
Discontinued
Annual number of pending appeals terminated during the course of proceedings before the court of second instance for formal reasons.
Other
Annual number of cases concluded with other rulings.
Pending
Annual number of cases not decided in a given year and remaining to be decided in subsequent years.
choose the category of data to be used for training & prediction
Press the suitable button to make your choice.
undecided appeals by years
Check out the data.
overview of rulings
The comparison of proportions of all types of rulings in the beginning and in the end of the period under analysis.
2011
2021
cases completed vs undecided
2011
Cases completed vs undecided against the total number of appeals brought in a particular year.
57274
47784
9482
cases completed vs undecided
2021
Cases completed vs undecided against the total number of appeals brought in a particular year.
112361
72697
39664
substantive decisions
2011
Cases in which the court ruled on their substance.
57274
31161
10592
3118
1164
46035
substantive decisions
2021
Cases in which the court ruled on their substance.
112361
41419
17291
2760
2052
63522
train AI and predict
The training lasts 200 epochs, i.e. the cycle of corrections of the model’s accuracy. The entire training should not take more than 5 minutes. The resulting prediction will be made automatically afterwards. Use the slider to select the hypothetical annual number of appeals brought in for which the prediction will occur.
AI training on your mobile
Because the computing workload I recommend not to train the model with phones, especially the older ones.
Number of epochs200
Appeals in: 83724
prediction result
After AI is trained you will see the result of its work.
created by Tadeusz Mięsowicz | 2022 | All right reserved