# Anomaly Detection
`Tutorials` are the best documentation — [`Anomaly Detection Tutorial`](https://colab.research.google.com/github/sovai-research/sovai-public/blob/main/notebooks/computational/Anomaly%20Detection.ipynb)
### Key Features
1. Multiple anomaly detection methods:
* Global anomalies: Identify outliers considering the entire dataset
* Local anomalies: Detect outliers within local neighborhoods
* Cluster anomalies: Find anomalies considering multi-dimensional data structure
2. Anomaly scoring: Compute anomaly scores for each data point
3. Feature-level anomaly analysis: Identify the most anomalous features for a given security
### Usage
Load the accounting factors data:
```python
import pandas as pd
df_factors = sov.data("factors/accounting", purge_cache=True)
df_last_3_years = df_factors.loc[(slice(None), slice(pd.Timestamp.now() - pd.DateOffset(years=3), None)), :]
df_last_3_years = df_last_3_years.percentile()
```
### Anomaly Detection
#### Compute Anomaly Scores
#### Local Anomalies
```python
df_local = df_last_3_years.anomalies("local", ticker="NVDA")
```
#### Global Anomalies
```python
df_global = df_last_3_years.anomalies("global", ticker="NVDA")
```
#### Cluster Anomalies
```python
df_cluster = df_last_3_years.anomalies("cluster", ticker="NVDA")
```
### Notes
* The module uses the `sovai` library for data loading and processing. Ensure you have the necessary permissions and valid authentication token.
* Anomaly detection methods can be applied to different time ranges and tickers. Adjust the parameters as needed for your analysis.
* The module provides flexibility in analyzing anomalies at both the overall and feature level. Experiment with different combinations of methods for comprehensive insights.
* When working with large datasets, be mindful of computational resources, especially when applying multiple anomaly detection methods or creating complex visualizations.
%20(1)%20(1)%20(1)%20(1).png?alt=media)
%20(1)%20(1)%20(1).png?alt=media)
%20(1)%20(1)%20(1).png?alt=media)
%20(1)%20(1).png?alt=media)