分析与 KPI 计算
使用分组聚合、透视和滚动计算月度队列留存率、产品级收入和滚动 30 天活跃用户数
分析与 KPI 计算 是 CoddyKit 上的免费 Pandas & NumPy Academy 课时。 这是第 3 节课，共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 Pandas & NumPy Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 Pandas & NumPy Academy 课程共包含 4 节课。
本课时的部分内容尚未翻译，以英文显示。
Loading Analysis-Ready Data
With cleaning and feature engineering complete, load the analysis_ready.parquet checkpoint and begin computing KPIs. This stage is pure analysis — no more data cleaning. Having a clean, validated checkpoint lets you iterate quickly on KPI definitions without re-running the time-consuming ingestion and cleaning stages. The three KPIs to compute: monthly cohort retention, product-level revenue by category, and rolling 30-day active users.
import pandas as pd
df = pd.read_parquet('output/analysis_ready.parquet')
print(f'Loaded: {df.shape}')
print('KPIs to compute:')
print(' 1. Monthly cohort retention matrix')
print(' 2. Category revenue and margin ranking')
print(' 3. Rolling 30-day active users per region')KPI 1: Monthly Revenue by Category
The first KPI is monthly revenue by category. Group by both year_month and category, sum revenue, and pivot to get categories as columns and months as rows. This pivot table is the basis of the trend line chart and the top-category bar chart in the report. Apply a 3-month rolling mean to smooth the trend and highlight seasonality.
import pandas as pd
df = pd.read_parquet('output/analysis_ready.parquet')
# Monthly revenue by category
monthly_cat = (
df.groupby(['year_month', 'category'])['revenue']
.sum()
.reset_index()
.pivot(index='year_month', columns='category', values='revenue')
.fillna(0)
.sort_index()
)
# 3-month rolling average
monthly_cat_smooth = monthly_cat.rolling(3, min_periods=1).mean()
print('Monthly revenue by category (head):')
print(monthly_cat.round(0).head())
print('Shape:', monthly_cat.shape)KPI 2: Top Regions by Total Revenue
Rank regions by total revenue across the entire period. Use groupby().agg() to compute multiple statistics simultaneously: total revenue, order count, unique customers, and average order value. Sort by total revenue descending and keep the top 10. Compute the revenue share of each region relative to the grand total — this reveals concentration risk (if 80% of revenue comes from one region, the business is geographically exposed).
import pandas as pd
df = pd.read_parquet('output/analysis_ready.parquet')
region_kpi = (
df.groupby('region')
.agg(
total_revenue=('revenue', 'sum'),
order_count=('order_id', 'nunique'),
unique_customers=('customer_id', 'nunique'),
avg_order_value=('revenue', 'mean')
)
.reset_index()
.sort_values('total_revenue', ascending=False)
)
total = region_kpi['total_revenue'].sum()
region_kpi['revenue_share'] = (region_kpi['total_revenue'] / total).round(3)
print(region_kpi.head(10).to_string(index=False))KPI 3: Cohort Retention Matrix
The cohort retention matrix shows what percentage of customers from each acquisition cohort made at least one purchase in each subsequent period. Compute it in three steps: (1) assign each order a 'period number' = number of months since the customer's cohort month; (2) groupby cohort and period number, count distinct customers; (3) divide by the cohort size (period 0 count) to get retention rates. Pivot to cohort × period format.
import pandas as pd
df = pd.read_parquet('output/analysis_ready.parquet')
# Compute months since cohort start
df['cohort_period'] = (
(df['order_date'].dt.to_period('M') -
pd.PeriodIndex(df['cohort_month'], freq='M'))
.apply(lambda x: x.n)
)
cohort_data = (
df.groupby(['cohort_month', 'cohort_period'])['customer_id']
.nunique()
.reset_index()
.rename(columns={'customer_id': 'customers'})
)
# Pivot: rows=cohort, columns=period
cohort_pivot = cohort_data.pivot(
index='cohort_month', columns='cohort_period', values='customers'
)
cohort_sizes = cohort_pivot[0]
retention = cohort_pivot.divide(cohort_sizes, axis=0).round(3)
print('Retention matrix (first 3 cohorts):')
print(retention.head(3))KPI 4: Rolling 30-Day Active Users
Daily Active Users (DAU) and rolling window variants are core product KPIs. Compute the number of distinct customers who purchased in any given 30-day window. Start by computing daily unique purchasers, set the date as the index, then apply rolling('30D').apply(lambda x: x.nunique()) — but note that rolling on a DatetimeIndex requires specific aggregation. An alternative is to use a set-based rolling unique count via a custom window function.
import pandas as pd
df = pd.read_parquet('output/analysis_ready.parquet')
# Daily unique purchasers
daily = (
df.groupby('order_date')['customer_id']
.nunique()
.reset_index()
.rename(columns={'customer_id': 'daily_unique_customers'})
.set_index('order_date')
.sort_index()
)
# 30-day rolling sum (approximation: sum of daily unique purchasers)
# True rolling unique requires custom logic
daily['rolling_30d'] = daily['daily_unique_customers'].rolling('30D').sum()
print('Rolling 30-day active customers:')
print(daily.tail())Product-Level Revenue Ranking
For product-level analysis, rank all product IDs by total revenue and compute their cumulative revenue share. This reveals whether revenue follows a Pareto distribution (a common finding: ~20% of products generate ~80% of revenue). Compute the cumulative revenue column with .cumsum() and find the cut-off product count where cumulative share exceeds 80%. This analysis guides decisions about product portfolio prioritisation.
import pandas as pd
df = pd.read_parquet('output/analysis_ready.parquet')
product_rev = (
df.groupby('product_id')['revenue']
.sum()
.sort_values(ascending=False)
.reset_index()
)
total_rev = product_rev['revenue'].sum()
product_rev['rev_share'] = product_rev['revenue'] / total_rev
product_rev['cumulative_share'] = product_rev['rev_share'].cumsum()
pct80_cutoff = (product_rev['cumulative_share'] <= 0.80).sum()
total_products = len(product_rev)
print(f'Top {pct80_cutoff} of {total_products} products ({pct80_cutoff/total_products:.0%}) '
f'generate 80% of revenue (Pareto)')
print(product_rev.head())Saving All KPI Tables
Save each KPI table to a named Parquet file in the output directory. Naming convention: kpi_{name}.parquet. This makes the visualisation stage straightforward — it reads exactly the pre-computed table it needs without recomputing anything. Also save a combined KPI summary to a CSV for easy sharing with non-technical stakeholders who may open it in Excel.
import pandas as pd
# Save all KPI tables
monthly_cat.to_parquet('output/kpi_monthly_category_revenue.parquet')
region_kpi.to_parquet('output/kpi_region_summary.parquet', index=False)
retention.to_parquet('output/kpi_cohort_retention.parquet')
daily.to_parquet('output/kpi_rolling_active_users.parquet')
product_rev.to_parquet('output/kpi_product_ranking.parquet', index=False)
# Also a CSV summary for sharing
region_kpi.to_csv('output/region_summary.csv', index=False)
print('All KPI tables saved:')
print(' kpi_monthly_category_revenue.parquet')
print(' kpi_region_summary.parquet')
print(' kpi_cohort_retention.parquet')
print(' kpi_rolling_active_users.parquet')
print(' kpi_product_ranking.parquet')Statistical Validation of KPIs
Before finalising KPIs, run statistical checks: test whether monthly revenue growth is statistically significant using a one-sample t-test against zero growth, and check whether region revenue differences are significant using one-way ANOVA. Reporting 'top region by revenue' means more when you can confirm that the difference is unlikely due to random variation. Append p-values to the KPI tables as metadata columns.
import pandas as pd
import numpy as np
from scipy import stats
df = pd.read_parquet('output/analysis_ready.parquet')
# Is revenue growth statistically significant?
monthly_total = df.groupby('year_month')['revenue'].sum().sort_index()
month_changes = monthly_total.diff().dropna()
stat, p = stats.ttest_1samp(month_changes, popmean=0)
print(f'Monthly growth test: mean={month_changes.mean():.0f}, p={p:.4f}')
print('Significant positive trend?', month_changes.mean() > 0 and p < 0.05)
# ANOVA: do regions differ significantly?
groups = [g['revenue'].values for _, g in df.groupby('region')]
f, p_anova = stats.f_oneway(*groups)
print(f'Region ANOVA: F={f:.3f}, p={p_anova:.4f}')Computing Month-over-Month Growth Rate
The month-over-month (MoM) growth rate is a fundamental business KPI: growth = (current - previous) / previous × 100. Use pct_change() for clean computation. A 3-month moving average of MoM growth reveals the underlying trend, smoothing out monthly volatility. Negative growth rates are just as important to highlight as positive ones — they signal periods requiring investigation.
import pandas as pd
df = pd.read_parquet('output/analysis_ready.parquet')
monthly_total = (
df.groupby('year_month')['revenue'].sum().sort_index()
)
growth = monthly_total.pct_change() * 100
trend = growth.rolling(3, min_periods=1).mean()
kpi_growth = pd.DataFrame({
'revenue': monthly_total,
'mom_growth_pct': growth.round(2),
'trend_3m_avg': trend.round(2)
})
print(kpi_growth.tail(6))
print(f'\nBest month: {monthly_total.idxmax()} (${monthly_total.max():,.0f})')
print(f'Worst month: {monthly_total.idxmin()} (${monthly_total.min():,.0f})')Automated Anomaly Detection
Flag anomalous months automatically: months where revenue is more than 2 standard deviations from the trailing 6-month mean. These are worth investigating — they may reflect genuine business events (a successful campaign, a data outage) or data quality issues (duplicate data loaded twice). Anomaly detection is a valuable addition to any automated pipeline dashboard and helps analysts focus their investigation on the most important time periods.
import pandas as pd
import numpy as np
df = pd.read_parquet('output/analysis_ready.parquet')
monthly = df.groupby('year_month')['revenue'].sum().sort_index()
rolling_mean = monthly.rolling(6, min_periods=3).mean()
rolling_std = monthly.rolling(6, min_periods=3).std()
anomalies = monthly[
(monthly > rolling_mean + 2 * rolling_std) |
(monthly < rolling_mean - 2 * rolling_std)
]
print('Anomalous months (>2 std from 6-month rolling mean):')
if len(anomalies) > 0:
print(anomalies.round(0))
else:
print('None detected')KPI Summary Table
Compile a single executive summary table combining the top-level values from all KPIs: total revenue, total orders, unique customers, average order value, best region, best category, and average 30-day cohort retention at month 3. This one-table summary is the first page of the report — it gives stakeholders an immediate headline view before they dive into charts and detailed tables.
import pandas as pd
df = pd.read_parquet('output/analysis_ready.parquet')
retention = pd.read_parquet('output/kpi_cohort_retention.parquet')
region_kpi = pd.read_parquet('output/kpi_region_summary.parquet')
summary = {
'Total Revenue': f'${df["revenue"].sum():,.0f}',
'Total Orders': f'{df["order_id"].nunique():,}',
'Unique Customers': f'{df["customer_id"].nunique():,}',
'Avg Order Value': f'${df["revenue"].mean():.2f}',
'Best Region': region_kpi.iloc[0]["region"],
'Best Category': df.groupby("category")["revenue"].sum().idxmax(),
'Avg Month-3 Retention': f'{retention.get(3, pd.Series([0])).mean():.1%}'
}
for k, v in summary.items():
print(f'{k:25s}: {v}')Quick Check
Test your understanding of Data Analysis concepts from this lesson.
Lesson Recap
In this lesson you learned: groupby + pivot creates the monthly category revenue matrix and cohort retention table, rolling() computes smoothed trends and 30-day active user approximations, and statistical validation (t-test for growth, ANOVA for region differences) adds rigor to the KPI story. Next up we visualise all results in a multi-panel figure and export the final report.
常见问题解答
「分析与 KPI 计算」课时是免费的吗？
是的 — 「分析与 KPI 计算」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 Pandas & NumPy Academy 课程的其余内容，请升级到 CoddyKit PRO。 Pandas & NumPy Academy 课程共包含 4 节课。
「分析与 KPI 计算」这节课中我会学到什么？
使用分组聚合、透视和滚动计算月度队列留存率、产品级收入和滚动 30 天活跃用户数 你通过在浏览器中直接运行的动手代码来练习 Pandas & NumPy Academy，全天候 AI 导师会在你学习这节课的过程中回答你的问题。
学习 Pandas & NumPy Academy 需要有经验吗？
无需任何先前经验。CoddyKit 上的 Pandas & NumPy Academy 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。 这是第 3 节课，共 4 节。
「分析与 KPI 计算」课时需要多长时间？
大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。
我能在这节 Pandas & NumPy Academy 课中编写并运行代码吗？
能。每节 Pandas & NumPy Academy 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。
此课程中的所有课时
- 项目设置与数据导入
- 数据清洗与特征工程
- 分析与 KPI 计算
- 最终可视化与报告导出