Feature Engineering for Time Series

# Import pandas and numpy
import pandas as pd
import numpy as np

# Create a daily energy demand DataFrame spanning 14 days across a month boundary
energy_df = pd.DataFrame({
    'date':   pd.date_range(start='2024-01-28', periods=14, freq='D'),  # 14 days including month boundary
    'demand': [310, 295, 280, 275, 320, 330, 315,                       # Jan 28 – Feb 3
               305, 290, 278, 272, 325, 335, 318]                       # Feb 4 – Feb 10 (MWh)
})

# --- Step 1: Extract raw calendar features from the datetime column ---
energy_df['year']        = energy_df['date'].dt.year          # calendar year
energy_df['month']       = energy_df['date'].dt.month         # 1=Jan, 12=Dec
energy_df['day']         = energy_df['date'].dt.day           # day of month (1–31)
energy_df['dayofweek']   = energy_df['date'].dt.dayofweek     # 0=Monday, 6=Sunday
energy_df['quarter']     = energy_df['date'].dt.quarter       # Q1=1, Q4=4
energy_df['weekofyear']  = energy_df['date'].dt.isocalendar().week.astype(int)  # ISO week number
energy_df['is_weekend']  = (energy_df['dayofweek'] >= 5).astype(int)            # 1 if Sat or Sun

# --- Step 2: Sine-cosine encoding for cyclical features ---
# Why: month 12 and month 1 are adjacent in time but 11 integers apart numerically
# Sine-cosine encoding places them on a circle so they are spatially close

# Month encoding: period = 12 months
energy_df['month_sin'] = np.sin(2 * np.pi * energy_df['month'] / 12)   # sine component
energy_df['month_cos'] = np.cos(2 * np.pi * energy_df['month'] / 12)   # cosine component

# Day-of-week encoding: period = 7 days
energy_df['dow_sin']   = np.sin(2 * np.pi * energy_df['dayofweek'] / 7)  # sine of day-of-week
energy_df['dow_cos']   = np.cos(2 * np.pi * energy_df['dayofweek'] / 7)  # cosine of day-of-week

# Round for display
energy_df = energy_df.round(3)

# Print calendar features and sine-cosine encoding
print("Calendar + cyclical features:")
print(energy_df[['date','month','dayofweek','is_weekend',
                  'month_sin','month_cos','dow_sin','dow_cos','demand']].to_string(index=False))

# Import pandas and numpy
import pandas as pd
import numpy as np

# Recreate the energy DataFrame with 14 rows
energy_df = pd.DataFrame({
    'date':   pd.date_range(start='2024-01-28', periods=14, freq='D'),
    'demand': [310, 295, 280, 275, 320, 330, 315,
               305, 290, 278, 272, 325, 335, 318]
}).sort_values('date').reset_index(drop=True)  # ensure chronological order

# --- Lag features ---
energy_df['lag_1'] = energy_df['demand'].shift(1)   # yesterday's demand
energy_df['lag_7'] = energy_df['demand'].shift(7)   # same day last week

# --- Rolling features ---
energy_df['roll_mean_3d'] = energy_df['demand'].rolling(window=3, min_periods=1).mean()  # 3-day avg
energy_df['roll_std_3d']  = energy_df['demand'].rolling(window=3, min_periods=1).std()   # 3-day volatility
energy_df['roll_mean_7d'] = energy_df['demand'].rolling(window=7, min_periods=1).mean()  # 7-day avg

# --- Day-over-day change ---
energy_df['delta_lag_1']  = energy_df['demand'] - energy_df['lag_1']  # raw change vs yesterday

# --- Trend indicator: is demand above its 7-day rolling average? ---
# +1 = above trend (rising), -1 = below trend (falling), 0 = exactly on trend (rare)
energy_df['trend_flag']   = np.sign(energy_df['demand'] - energy_df['roll_mean_7d']).astype(int)

# --- Demand deviation from rolling mean (z-score style) ---
energy_df['demand_zscore'] = (
    (energy_df['demand'] - energy_df['roll_mean_7d']) /
    (energy_df['roll_std_3d'] + 1e-9)
)

# Round for clean display
energy_df = energy_df.round(2)

# Print the enriched DataFrame
print(energy_df[['date','demand','lag_1','lag_7','roll_mean_3d',
                  'roll_mean_7d','delta_lag_1','trend_flag','demand_zscore']].to_string(index=False))

# Import pandas and numpy
import pandas as pd
import numpy as np

# Set a fixed reference date — the "today" from the model's perspective
# In production this would be datetime.today() or the pipeline run date
reference_date = pd.Timestamp('2024-03-01')  # snapshot date: March 1, 2024

# Create a customer activity DataFrame — 10 rows
customers_df = pd.DataFrame({
    'customer_id':    [1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
    'last_login':     pd.to_datetime([
        '2024-02-29','2024-01-15','2024-02-25','2023-12-01','2024-02-28',
        '2024-01-30','2023-11-15','2024-02-20','2024-02-01','2023-10-10'
    ]),   # date of most recent login
    'last_purchase':  pd.to_datetime([
        '2024-02-20','2024-01-10','2024-02-10','2023-11-15','2024-02-25',
        '2024-01-05','2023-10-20','2024-01-28','2024-01-20','2023-09-01'
    ]),   # date of most recent purchase
    'signup_date':    pd.to_datetime([
        '2022-06-01','2021-03-15','2023-01-10','2020-09-20','2022-11-05',
        '2021-07-22','2020-04-18','2023-05-30','2022-08-14','2019-12-01'
    ]),   # date customer first signed up
    'will_churn':     [0, 1, 0, 1, 0, 1, 1, 0, 0, 1]  # target: 1=will churn in 30 days
})

# --- Time since last login (days) ---
customers_df['days_since_login']    = (reference_date - customers_df['last_login']).dt.days

# --- Time since last purchase (days) ---
customers_df['days_since_purchase'] = (reference_date - customers_df['last_purchase']).dt.days

# --- Customer tenure: days since signup ---
customers_df['tenure_days']         = (reference_date - customers_df['signup_date']).dt.days

# --- Gap between last login and last purchase (days) ---
# Large gap = customer logs in but doesn't buy — a potential disengagement signal
customers_df['login_purchase_gap']  = (
    customers_df['last_login'] - customers_df['last_purchase']
).dt.days

# --- Recency score: inverse of days_since_login — higher = more recent ---
# Adding 1 to avoid division by zero for a same-day login
customers_df['recency_score']       = 1 / (customers_df['days_since_login'] + 1)

# Check class separation for each time-based feature
print("Class separation — time features vs churn:\n")
sep = customers_df.groupby('will_churn')[
    ['days_since_login','days_since_purchase','tenure_days','login_purchase_gap']
].mean().round(1)
print(sep.to_string())

# Print full DataFrame
print("\nFull customer time-feature DataFrame:")
print(customers_df[['customer_id','days_since_login','days_since_purchase',
                     'tenure_days','login_purchase_gap','recency_score','will_churn']].round(4).to_string(index=False))