import pandas as pd

daily = pd.read_csv("daily_bookings.csv", parse_dates=["date"]).set_index("date")
# Seasonal naive baseline: predict each day equal to the same day last year
daily["forecast"] = daily["bookings"].shift(364)      # 52 weeks keeps the weekday aligned
test = daily.loc["2026-01-01":].dropna()
mae = (test["bookings"] - test["forecast"]).abs().mean()
print(f"Seasonal naive MAE: {mae:.1f} bookings/day")
# Any fancier model should beat this baseline

Cancellation model: allowed vs leaking features
Allowed (known at booking time):
  lead_time_days, nights, room_type, channel, payment_type, past_cancellations
Leaking (only known later — exclude):
  days_until_cancellation, refund_amount, modification_count_after_booking
Rule: only use information available at the moment you make the prediction.

import pandas as pd
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import KMeans

cust = pd.read_csv("customers.csv")   # trips_per_year, avg_spend, avg_lead_days, family_trips_share
features = ["trips_per_year", "avg_spend", "avg_lead_days", "family_trips_share"]
X = StandardScaler().fit_transform(cust[features])
cust["segment"] = KMeans(n_clusters=4, n_init=10, random_state=7).fit_predict(X)
print(cust.groupby("segment")[features].mean().round(1))
# Name segments after inspecting their averages, e.g. 'frequent business', 'family planners'

import pandas as pd

reviews = pd.read_csv("reviews.csv")   # text, rating
themes = {"cleanliness": ["clean", "dirty"], "staff": ["staff", "reception"],
          "location": ["location", "near"], "value": ["price", "value"]}
text = reviews["text"].str.lower()
for theme, words in themes.items():
    mask = text.str.contains("|".join(words))
    print(f"{theme:12} mentions={mask.sum():4}  avg_rating={reviews.loc[mask, 'rating'].mean():.2f}")

from statsmodels.stats.proportion import proportions_ztest, confint_proportions_2indep

bookings = [612, 668]        # control, variant
sessions = [24000, 24100]
stat, p = proportions_ztest(bookings, sessions)
low, high = confint_proportions_2indep(bookings[1], sessions[1], bookings[0], sessions[0])
print(f"p-value: {p:.3f}")
print(f"Difference in conversion (variant - control): {low:.4f} to {high:.4f}")

Experiment brief
Change: show 'free cancellation' badge on search results
Hypothesis: clearer policy increases booking conversion
Primary metric: bookings / search sessions
Guardrails: cancellation rate, page load time
Duration: two full weeks (avoid starting before a holiday)
Decision rule: ship if conversion improves and guardrails hold

Travel data science projects (public datasets)
1. Hotel booking cancellation prediction with leakage-free features
2. Monthly booking forecast with holiday features vs seasonal naive baseline
3. Traveller segmentation with named, actionable segments
4. Review theme analysis linking themes to ratings
Each: problem, data, method, results, recommendation

Practise these case prompts
- 'Bookings for hill stations dropped this month. How would you investigate?'
- 'How would you predict which bookings will be cancelled?'
- 'Design an experiment for a new search filter.'
Structure: clarify → hypotheses → data → analysis → recommendation

One-page summary for a revenue team
Question: ______________
Key finding: ______________
Confidence: high / medium / low and why
Recommendation: ______________
What to watch next: ______________