The code for this project can be found at my GitHub.

该项目的代码可以在我的GitHub上找到 。

介绍 (Introduction)

The goal of this project was to analyse historic stock/inventory data to decide how much stock of each item a retailer should hold in the future. When deciding which historic data to use I came across a great blog post by Nhan Tran on Towards Data Science. This post provided me with the data set I was looking for. The best thing about this post was that it gave a brief outline of a project, but no code. This meant I had to write all the code myself from scratch (which is good practice), yet I could check my answers at each stage using the landmarks in Nhan Tran’s blog post.

该项目的目的是分析历史库存/库存数据，以确定零售商将来应持有的每件商品多少库存。 在决定使用哪些历史数据时，我遇到了Nhan Tran撰写的一篇很棒的博客文章 ，名为Towards Data Science 。 这篇文章为我提供了我想要的数据集 。 关于这篇文章的最好的事情是它给出了一个项目的简要概述，但是没有代码。 这意味着我必须自己重新编写所有代码(这是一种很好的做法)，但是我可以使用Nhan Tran 博客文章中的地标在每个阶段检查答案。

I completed the project using Python in a Jupyter Notebook.

我在Jupyter Notebook中使用Python完成了该项目 。

技术目标 (Technical Goal)

To give (with 95% confidence) a lower and upper bound for how many pairs of Men’s shoes, of each size, a shop in the United States should stock for any given month.

为了给(具有95％的置信度)上下限，确定每个给定月份在美国的商店应存多少种每种尺寸的男鞋。

设置笔记本并导入数据 (Setting Up the Notebook and Importing the Data)

First I read through the project and decided which libraries I would need to use. I then imported those libraries into my notebook:

首先，我通读该项目，并确定需要使用哪些库。 然后，我将这些库导入到笔记本中：

Note: some of these libraries I didn’t foresee using and I went back and added them as and when I needed them.

注意 ：其中一些我没有预见到的库，我回去并在需要时添加它们。

Next I downloaded the data set and imported it into my notebook using read_csv() from the pandas library:

接下来，我从pandas库中使用read_csv()下载了数据集并将其导入到笔记本中：

sales = pd.read_csv("Al-Bundy_raw-data.csv")

检查和清理数据 (Inspecting and Cleaning the Data)

Before beginning the project I inspected the DataFrame.

在开始项目之前，我检查了DataFrame。

Each row on the data set represent a pair of shoes that had been sold.

数据集的每一行代表一双已售出的鞋子。

I then inspected the data types of each column to ensure they made sense.

然后，我检查了每一列的数据类型，以确保它们有意义。

I then changed the 'Date' datatype from object to datetime using the following code:

然后我改变了'Date'数据类型从object到datetime使用以下代码：

sales['Date'] = pd.to_datetime(sales['Date'])

Based on the specific goal of this project I then decided to drop any columns that would not be relevant during this analysis. I dropped the European and UK size equivalents as only one size type was needed, and since I was focusing on US stores in particular I decided to keep US sizes to analyse. I dropped the Date column as I already had the month and year data in separate columns and did not need to be more specific than this. I also dropped the Product ID, Invoice Number, Discount and Sale Price as they were either irrelevant or their effect on stock levels was beyond the scope of this project. I achieved this using the following code:

基于此项目的特定目标，我然后决定删除此分析期间不相关的任何列。 因为只需要一种尺码类型，所以我放弃了欧洲和英国的尺码对应关系，并且由于我特别关注美国商店，因此我决定保留美国尺码进行分析。 我删除了Date列，因为我已经在不同的列中包含了月份和年份数据，并且不需要比此更具体。 我还删除了产品ID，发票编号，折扣和销售价格，因为它们无关紧要或它们对库存水平的影响超出了此项目的范围。 我使用以下代码实现了这一点：

sales_mini = sales.drop(['InvoiceNo', 'ProductID','Size (Europe)', 'Size (UK)', 'UnitPrice', 'Discount', 'Date'], axis=1)

This left me with a more streamlined and easier to read DataFrame to analyse.

这为我提供了更简化和更易于读取的DataFrame进行分析。

数据分析与可视化 (Data Analysis and Visualisation)

I now decided to concentrate on an even smaller set of data, although this may have sacrificed accuracy by using a smaller data set, it also may have gained accuracy by using the most relevant data available.

我现在决定集中精力于更小的数据集，尽管使用较小的数据集可能会牺牲准确性，但使用可用的最相关数据也可能会提高准确性。

To do this I first created a new DataFrame of only Male shoes, using the following code:

为此，我首先使用以下代码创建了一个仅男鞋的新数据框：

sales_mini_male = sales_mini[sales_mini['Gender'] == 'Male']

I then selected all of the rows where the shoes were sold in the United States in 2016, in order to make the data as relevant to our goal as possible. I achieved this with the following code:

然后，我选择了2016年在美国销售鞋子的所有行，以使数据尽可能与我们的目标相关。 我通过以下代码实现了这一点：

male_2016 = sales_mini_male[sales_mini_male['Year']==2016]
male_us_2016 = male_2016[male_2016['Country']=='United States']

The above DataFrame doesn’t make it clear which shoe sizes were the most frequent. To find out which shoe sizes were, I created a pivot table, analysing how many shoes of each size were sold each month (since people may be more likely to buy shoes in some months that others). I achieved this using the following code:

上面的DataFrame并不清楚哪个鞋号是最常见的。 为了找出哪种尺寸的鞋，我创建了一个数据透视表，分析了每个月售出每种尺寸的鞋的数量(因为人们可能会在几个月内购买其他鞋的可能性更高)。 我使用以下代码实现了这一点：

male_us_2016_by_month = pd.pivot_table(male_us_2016, values='Country', index=['Size (US)'], columns=['Month'], fill_value=0, aggfunc=len)

Albeit more useful, the above table is still difficult to read. To solve this problem I imported the seaborn library, and displayed the pivot table as a heat map.

上表虽然更有用，但仍然很难阅读。 为了解决此问题，我导入了seaborn库，并将数据透视表显示为热图。

plt.figure(figsize=(16, 6))
male_us_2016_by_month_heatmap = sns.heatmap(male_us_2016_by_month, annot=True, fmt='g', cmap='Blues')

This heat map indicates that demand for shoes, across different sizes is likely to be a normal distribution. This makes sense when thinking logically, as few people have extremely small or extremely large feet, but many have a shoe size somewhere in between. I then illustrated this even clearer by plotting the total yearly demand for each size in a bar chart.

该热图表明，不同尺寸的鞋子需求可能是正态分布。 从逻辑上考虑时，这是有道理的，因为很少有人脚很小或太大，但是很多人的鞋子介于两者之间。 然后，我通过在条形图中绘制每种尺寸的年度总需求来说明这一点。

male_us_2016_by_month_with_total = pd.pivot_table(male_us_2016, values='Country', index=['Size (US)'], columns=['Month'], fill_value=0, margins=True, margins_name='Total', aggfunc=len)
male_us_2016_by_month_with_total_right = male_us_2016_by_month_with_total.iloc[:-1, :]
male_us_2016_by_month_with_total_right = male_us_2016_by_month_with_total_right.reset_index()
male_us_2016_total_by_size = male_us_2016_by_month_with_total_right[['Size (US)', 'Total']]
male_us_2016_by_size_plot = male_us_2016_total_by_size.plot.bar(x='Size (US)',y='Total', legend=False)
male_us_2016_by_size_plot.set_ylabel("Frequency")

Although this analysis would be useful to a retailer, this is just an overview of what happened in 2016.

尽管此分析对零售商有用，但这只是2016年情况的概述。

学生考试 (Student’s T-test)

To give a prediction on future levels of demand, with a given degree of confidence, I performed a statistical test on the data. Since the remaining relevant data was a small data set, and that it closely represented a normal distribution, I decided a Student’s T-test would be the most relevant.

为了以给定的可信度预测未来的需求水平，我对数据进行了统计检验。 由于剩余的相关数据是一个很小的数据集，并且它紧密地代表了正态分布，因此我认为学生的T检验将是最相关的。

First I found the t-value to be used in a 2-tailed test finding 95% confidence intervals (note: 0.5 had to be divided by 2 to get 0.025 as the test is 2-tailed).

首先，我发现t值用于2尾检验中，发现置信区间为95％(注意：由于2尾检验，必须将0.5除以2才能得到0.025)。

t_value = stats.t.ppf(1-0.025,11)

I then used this t-value to calculate the ‘Margin Error’ and displayed this along with other useful aggregates in the following table using the following code:

然后，我使用此t值来计算“边距误差”，并使用以下代码在下表中将其与其他有用的汇总一起显示：

male_us_2016_agg = pd.DataFrame()
male_us_2016_agg['Size (US)'] = male_us_2016_by_month_with_total['Size (US)']
male_us_2016_agg['Mean'] = male_us_2016_by_month_with_total.mean(1)
male_us_2016_agg['Standard Error'] = male_us_2016_by_month_with_total.sem(1)
male_us_2016_agg['Margin Error'] = male_us_2016_agg['Standard Error'] * t_value
male_us_2016_agg['95% CI Lower Bound'] = male_us_2016_agg['Mean'] - male_us_2016_agg['Margin Error']
male_us_2016_agg['95% CI Upper Bound'] = male_us_2016_agg['Mean'] + male_us_2016_agg['Margin Error']

零售商友好输出 (Retailer Friendly Output)

I then decided to re-present my data in a more easy to understand format. I made sure python did as much of the work for me here as possible to ensure my code was efficient and easy to replicate and scale.

然后，我决定以一种更易于理解的格式重新呈现我的数据。 我确保python在这里为我做了尽可能多的工作，以确保我的代码高效且易于复制和扩展。

conclusion = pd.DataFrame()
conclusion['Size (US)'] = male_us_2016_agg['Size (US)']
conclusion['Lower Bound'] = male_us_2016_agg['95% CI Lower Bound'].apply(np.ceil)
conclusion['Lower Bound'] = conclusion['Lower Bound'].astype(int)
conclusion['Upper Bound'] = male_us_2016_agg['95% CI Upper Bound'].apply(np.floor)
conclusion['Upper Bound'] = conclusion['Upper Bound'].astype(int)
conclusion['Conclusion'] = np.where(conclusion['Size (US)'] == 'Total', 'Based on data from 2016, we would expect, with 95% confidence, to sell atleast ' + conclusion['Lower Bound'].astype(str) + ' pair(s), and upto ' + conclusion['Upper Bound'].astype(str) + ' pair(s) of shoes in a US store each month.', 'Based on data from 2016, we would expect, with 95% confidence, to sell atleast ' + conclusion['Lower Bound'].astype(str) + ' pair(s), and upto ' + conclusion['Upper Bound'].astype(str) + ' pair(s) of size ' + conclusion['Size (US)'].astype(str) + ' shoes in a US store each month.')
pd.set_option('display.max_colwidth',200)

可能进一步分析 (Possibly Further Analysis)

The same analysis above can all be completed for different 'Gender', 'Country' and 'Year' values. In total this would produce 2 x 4 x 3 = 24 different sets of bounds to guide the retailer. These bounds could be used to guide a retailer in each specific circumstance. Alternatively, if these results don't differ much we may use this as a reason to use a larger data set. For example, if our bounds don't change much for each 'Year', we may want to use the 'Male' 'United States' data from all years to get a more accurate result.

对于不同的'Gender' ， 'Country'和'Year'值，都可以完成以上相同的分析。 总体而言，这将产生2 x 4 x 3 = 24组不同的边界以指导零售商。 这些界限可以用来指导零售商在每种特定情况下。 另外，如果这些结果相差不大，我们可以以此为理由使用较大的数据集。 例如，如果每个'Year'界限变化不大，我们可能希望使用所有年份的'Male' 'United States'数据来获得更准确的结果。

Thanks for reading and thanks to Nhan Tran for the original post that guided this project.

感谢您的阅读，也感谢Nhan Tran提供了指导该项目的原始文章。

The code for this project can be found at my GitHub.

该项目的代码可以在我的GitHub上找到 。