Project — Property Price Predictor

• Build a regression pipeline for mixed-type property data.
• Evaluate with RMSE and R² (cross-validated).
• Engineer price-relevant features (price per sqft, age).
• Present predictions and uncertainty honestly.

If you don't have a real property CSV, generate a realistic synthetic one so the lesson runs anywhere:

import numpy as np, pandas as pd
rng = np.random.default_rng(0)
n = 600
df = pd.DataFrame({
    "sqft":     rng.integers(600, 3000, n),
    "rooms":    rng.integers(1, 6, n),
    "age":      rng.integers(0, 40, n),
    "type":     rng.choice(["apartment", "terrace", "flowerlow"], n),
    "city":     rng.choice(["Tokyo", "Cairo", "JB", "Lima"], n),
})
base = (df["sqft"] * 350 + df["rooms"] * 25000
        - df["age"] * 4000
        + df["type"].map({"apartment": 0, "terrace": 80000, "flowerlow": 250000})
        + df["city"].map({"Tokyo": 200000, "Cairo": 120000, "JB": 60000, "Lima": 0}))
df["price"] = (base + rng.normal(0, 50000, n)).round(-3).clip(lower=80000)
df.to_csv("property.csv", index=False)

The pieces (same as Titanic, Regressor at the end)

from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.ensemble import RandomForestRegressor

num_cols = ["sqft", "rooms", "age"]
cat_cols = ["type", "city"]

pre = ColumnTransformer([
    ("num", StandardScaler(), num_cols),
    ("cat", OneHotEncoder(handle_unknown="ignore"), cat_cols),
])
model = Pipeline([("prep", pre),
                  ("reg", RandomForestRegressor(n_estimators=300, random_state=0))])

Metrics for regression

from sklearn.model_selection import cross_val_score
r2  = cross_val_score(model, X, y, cv=5, scoring="r2").mean()
rmse = -cross_val_score(model, X, y, cv=5,
                        scoring="neg_root_mean_squared_error").mean()
print(f"R²: {r2:.3f}   RMSE: $ {rmse:,.0f}")

scikit-learn maximises scores, so error metrics are "negative" — flip the sign back. RMSE in $ is interpretable: "typical error of ±RMSE".

Feature engineering ideas

X["price_per_room_proxy"] = X["sqft"] / X["rooms"]
# age buckets, is_new, etc.

# property.py — predict & present
import pandas as pd
from sklearn.model_selection import train_test_split, cross_val_score
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import r2_score, mean_absolute_error, root_mean_squared_error
import matplotlib.pyplot as plt

df = pd.read_csv("property.csv")
y = df.pop("price")
X = df

pre = ColumnTransformer([
    ("num", StandardScaler(), ["sqft", "rooms", "age"]),
    ("cat", OneHotEncoder(handle_unknown="ignore"), ["type", "city"]),
])
model = Pipeline([("prep", pre),
                  ("reg", RandomForestRegressor(n_estimators=300, random_state=0))])

r2 = cross_val_score(model, X, y, cv=5, scoring="r2").mean()
print(f"CV R²: {r2:.3f}")

Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.2, random_state=0)
model.fit(Xtr, ytr)
pred = model.predict(Xte)
print(f"test RMSE : $ {root_mean_squared_error(yte, pred):,.0f}")
print(f"test MAE  : $ {mean_absolute_error(yte, pred):,.0f}")

# predicted vs actual
plt.figure(figsize=(6, 6))
plt.scatter(yte, pred, alpha=0.4, s=12)
lim = [yte.min(), yte.max()]
plt.plot(lim, lim, "r--")
plt.xlabel("actual ($)"); plt.ylabel("predicted ($)")
plt.title("Property price: predicted vs actual")
plt.tight_layout(); plt.savefig("property_pred.png", dpi=150)

# predict a new property, with an error band
new = pd.DataFrame([{"sqft": 1500, "rooms": 3, "age": 5,
                     "type": "terrace", "city": "Tokyo"}])
p = model.predict(new)[0]
rmse = root_mean_squared_error(yte, pred)
print(f"\nestimate: $ {p:,.0f}  (± $ {rmse:,.0f})")

Sample output

CV R²: 0.942
test RMSE : $58,400
test MAE  : $44,100

estimate: $742,000  (± $58,400)

Read the diff

The estimate comes with an honest ± band derived from the test RMSE — never quote a price to the nearest dollar. High R² here is partly because our synthetic data is fairly linear; real property data is noisier (location micro-effects, renovation, timing), so expect lower R² and wider bands in the wild.

Regression = same pipeline, Regressor at the end, RMSE/R² instead of accuracy. Always present predictions with an error band. Save the whole pipeline with joblib so prep + model deploy as one. You now have two complete projects — and the classic-ML toolkit is done. Next: neural networks.