Crazy Simple Anomaly Detection for Customer Success

1_Crazy Simple Anomaly Detection for Customer Success

# Crazy Simple Anomaly Detection for Customer Success

anomalies.py

# Create Pandas DataFrame
temp = pd.concat(
    [
        x,
        y,
        rolling_mad,
        rolling_mad_upper,
        rolling_mad_lower,
    ],
    axis=1,
)

# Rename columns
temp.columns = ["timeline", "values", "mad", "upper_mad", "lower_mad"]

# Drop initial values (outside rolling window)
temp.dropna(subset=["mad"], inplace=True)

# Determine anomalies
temp["is_anomaly"] = ~(
    temp["values"].between(temp["lower_mad"], temp["upper_mad"])
)
temp["anomaly_status"] = temp["is_anomaly"].apply(
    lambda v: "ANOMALY" if v else "NOT ANOMALY"
)

bounds.py

threshold_ratio = 0.5

rolling_mad_upper = rolling_mad * (1 + threshold_ratio)
rolling_mad_lower = rolling_mad * (1 - threshold_ratio)

mad.py

mad = lambda x: np.median(np.fabs(x - np.median(x)))

rolling_mad.py

# We'll use a 7-day window
rolling_window = 7

rolling_mad = y.rolling(window=rolling_window, center=False).median() 
	+ y.rolling(window=rolling_window, center=False).apply(mad, raw=True)

select_data.py

X = df["day"]
Y = df["customer_3"]