FinTools.md

Financial analytics tools for merchants

Data Collected

Customers

By storing and analysing the payment card data, we can derive the following:

• New vs returning customers
• Geographical origin

Products

Tracking sales gives us:

• Peak times
• Product performance
• Store vs store performance
• Sales trends
  • hourly (coffees are mostly sold at 8-9am and 6-7pm)
  • daily (Sunday sales for trousers are highest, tshirts mostly sell on Wed)
  • weekly
  • seasonal

Transactions

Looking at the transactions for a given time period, we can obtain:

• Avg. spent
• Card type
• Customer spending pattern

Predictions

Feeding the aforementioned data points, merchants can derive valuable data to optimise their CAC, loyalty programs and sales strategies.

Customers

• Loyalty campaigns: Use time-based insights to run time-sensitive promotions. For example, a “Happy Hour” discount during a known dull hour can attract more customers.
• Customer engagement: Tailor ads based on purchasing trends—promote items that have higher profit margins or that pair well together.

Inventory

• Efficient stock management - predict when an item will run out of stock

Promotions

• Evaluate efficacy of promotions respective to the season and the customer persona targeted.

Models

• ARIMA and Holt-Winters' models: predict inventory depletion, cashflow for the week or month, avg. transaction size
• Customer segmentation and Cohort analysis: separate customers to track favourite items and spending patterns
• Prophet (by Fb) and Random forest models: time series analysis tools, good for stock management and cashflow predictions
• Local Linear trend model: identify peaks in non-homogeneous datasets

$$ y_t = \mu_t + \epsilon_t $$

where

$$ \mu_{t+1} = \mu_t + \delta_t + \eta_t $$

and

$$ \delta_{t+1} = \delta_t + \zeta_t $$

All of these are within the variables $(0, \sigma^2)$ interval.

To the base model we can add a seasonality component, depending if the dataset is statistically significant in number (> 5) and the variance is reasonable according to a T-test.

$$ s_{t+1} = -\sum_{j=0}^{S-2} s_{t-j} + \omega_t, \quad \omega_t \sim N(0, \sigma_\omega^2) $$

• General GARCH model:

$$ \sigma_t^2 = \omega + \sum_{i=1}^{q} \alpha_i \epsilon_{t-i}^2 + \sum_{j=1}^{p} \beta_j \sigma_{t-j}^2 $$

Variance considerations as above.