How Companies Use Data Science to Solve Real Problems

import pandas as pd

# daily_sales.csv: date, units
df = pd.read_csv("daily_sales.csv", parse_dates=["date"]).set_index("date")

# Baseline 1: same day last week (seasonal naive)
df["naive_7"] = df["units"].shift(7)

# Baseline 2: average of the last 4 same weekdays
df["avg_4wk"] = (df["units"].shift(7) + df["units"].shift(14) +
                 df["units"].shift(21) + df["units"].shift(28)) / 4

test = df.loc["2026-09-01":].dropna()
for col in ["naive_7", "avg_4wk"]:
    mae = (test["units"] - test[col]).abs().mean()
    print(col, "MAE:", round(mae, 1))

from sklearn.ensemble import GradientBoostingRegressor
from sklearn.metrics import mean_absolute_error

# Continuing with df from the baseline tab
df["dow"] = df.index.dayofweek
df["month"] = df.index.month
df["lag_7"] = df["units"].shift(7)
df["lag_14"] = df["units"].shift(14)
data = df.dropna()

features = ["dow", "month", "lag_7", "lag_14"]
train = data.loc[:"2026-08-31"]
test = data.loc["2026-09-01":]

model = GradientBoostingRegressor(random_state=0)
model.fit(train[features], train["units"])
print("model MAE:", round(mean_absolute_error(test["units"], model.predict(test[features])), 1))
# Keep the model only if it clearly beats the baselines.

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

# customers.csv: customer_id, recency_days, frequency, monetary
rfm = pd.read_csv("customers.csv")
X = StandardScaler().fit_transform(rfm[["recency_days", "frequency", "monetary"]])

rfm["segment"] = KMeans(n_clusters=4, n_init=10, random_state=0).fit_predict(X)
print(rfm.groupby("segment")[["recency_days", "frequency", "monetary"]]
         .mean().round(1))
# Name segments by behaviour, e.g. 'loyal', 'at risk', 'new', 'occasional'

import pandas as pd
from sklearn.metrics.pairwise import cosine_similarity

# ratings.csv: user_id, item_id, rating
ratings = pd.read_csv("ratings.csv")
matrix = ratings.pivot_table(index="user_id", columns="item_id",
                             values="rating", fill_value=0)

# Item-to-item similarity
sim = pd.DataFrame(cosine_similarity(matrix.T),
                   index=matrix.columns, columns=matrix.columns)

def similar_items(item, n=5):
    return sim[item].drop(item).sort_values(ascending=False).head(n)

print(similar_items(matrix.columns[0]))

import numpy as np
from sklearn.ensemble import IsolationForest

rng = np.random.default_rng(0)
# Normal transactions: amount and hour of day
normal = np.column_stack([rng.normal(1500, 400, 1000), rng.normal(14, 3, 1000)])
# A few unusual ones: very large amounts at odd hours
odd = np.array([[48000, 3], [52000, 2], [39000, 4]])
X = np.vstack([normal, odd])

model = IsolationForest(contamination=0.005, random_state=0).fit(X)
flags = model.predict(X)              # -1 = anomaly
print("flagged rows:", np.where(flags == -1)[0][-5:])

Fraud model in practice (typical workflow)

1. Score each transaction in real time
2. Low score    -> approve automatically
3. Medium score -> extra verification (OTP, confirmation)
4. High score   -> hold and send to a human reviewer
5. Reviewer decision is recorded as a new label
6. Model is retrained regularly on new labels
7. Monitor: fraud caught, false alarms, customer complaints

Thresholds are business decisions, not only model settings.

Data science project charter

Business problem: ___________________________
Decision it will support: ____________________
Owner (business): ___________________________
Success metric: _____________________________
Current approach / baseline: _________________
Data available: ______________________________
Constraints: privacy, latency, explainability
Risks: _______________________________________
Timeline and checkpoints: ____________________

How will we measure real impact?

Before launch:
- Offline metric vs baseline: ________

After launch:
- A/B test or phased rollout?  ________
- Business metric to track: ________
- Guardrail metrics: ________
- Review date: ________

If impact is not visible, what will we change?
________________________________________

Turn a public dataset into a business case

1. Choose a domain (retail, transport, health, finance)
2. Write the business question as a manager would
3. Define the decision and success metric
4. Build a baseline, then a model
5. Translate results into an action
6. Estimate impact qualitatively (avoid made-up numbers)
7. List risks and next steps