Smart Inventory Advisor for Retail Sales (Invense)

This repository contains a machine learning model trained to predict daily sales for retail products and provide actionable inventory recommendations. The project was developed as part of the Datathon Nexus Crew.

Model Details

Model Description

The core of this project is a Random Forest Regressor (scikit-learn) trained to forecast the number of units sold (Sales) for a given product based on a variety of features. The primary output is a "Store Owner's Action Plan" that identifies products at risk of stocking out and recommends a reorder quantity.

Developed by: jerewy (NexusCrew)
Model type: RandomForestRegressor
Language(s) (NLP): en
License: MIT
Finetuned from model [optional]: This model was trained from scratch.

Model Sources [optional]

Repository: https://github.com/jerewy/datathon_nexus_crew

Uses

Direct Use

This model is intended to be used as a decision-support tool for small to medium-sized retail business owners. It helps answer two key questions:

Which products should I focus on restocking right now?
How many units of each product should I order?

The primary function, get_smart_inventory_recommendations in the notebook, automates this process.

Out-of-Scope Use

This model is not suitable for real-time stock market prediction or long-term (multi-year) financial forecasting. Its predictions are based on the patterns in the provided dataset and may not be accurate if underlying market conditions change drastically. The model should not be used for making automated financial decisions without human oversight.

Bias, Risks, and Limitations

The main limitation of this model is the "predictability ceiling" of the source data. The final model achieves a Testing R² score of 0.34, which means that while it successfully captures the predictable patterns, about 66% of the variance in daily sales is due to unpredictable, random factors not present in the data.

Data Dependency: The model's performance is entirely dependent on the quality and patterns of the training data. It assumes future trends will resemble historical ones.
Synthetic Data: The model was trained on a synthetic dataset. Performance on real-world, noisy retail data will likely differ and may require further tuning.

Recommendations

Users should be aware that the model provides a forecast, not a guarantee. The "Recommended Reorder Qty" should be treated as a strong, data-driven suggestion, but business owners should still apply their own domain knowledge, especially when considering external factors not included in the dataset (e.g., upcoming local events, new competitors).

How to Get Started with the Model

Use the code below to load the saved model and preprocessors to make predictions on new data.

import pickle
import pandas as pd

# Load the trained model, scaler, and encoders
model = pickle.load(open('sales_model.pkl', 'rb'))
scaler = pickle.load(open('scaler.pkl', 'rb'))
label_encoders = pickle.load(open('label_encoders.pkl', 'rb'))

# --- Example: Prepare a single row of new data ---
# NOTE: This must have the same structure as the training data
new_data = pd.DataFrame([{
    'Inventory': 200, 'Orders': 50, 'Price': 35.0, 'Discount': 10, 
    'Competitor Price': 33.0, 'Promotion': 1, 'Category': 'Groceries', 
    'Region': 'North', 'Weather': 'Sunny', 'Seasonality': 'Spring',
    'DayOfWeek': 2, 'Month': 4, 'Day': 15
}])

# Apply the same preprocessing
for col in ['Category', 'Region', 'Weather', 'Seasonality']:
    new_data[col] = label_encoders[col].transform(new_data[col])

# Scale the features
new_data_scaled = scaler.transform(new_data)

# Make a prediction
predicted_sales = model.predict(new_data_scaled)
print(f"Predicted Sales: {predicted_sales[0]:.2f} units")

Training Details

Training Data

The model was trained on the "Retail Store Inventory" dataset, which contains over 73,000 daily records across multiple stores and products. The data is synthetic but realistically models retail sales patterns.

Training Procedure

Preprocessing

The training data was preprocessed as follows:

Feature Engineering: DayOfWeek, Month, and Day were extracted from the Date column.
Label Encoding: Categorical features (Category, Region, Weather, Seasonality) were converted to numerical values.
Standard Scaling: All numerical features were scaled to have a mean of 0 and a standard deviation of 1.

Training Hyperparameters

The model was tuned using RandomizedSearchCV to find the optimal settings. The best-performing hyperparameters were:

n_estimators: 200
min_samples_split: 10
min_samples_leaf: 4
max_features: 'sqrt'
max_depth: 20

Evaluation

Testing Data, Factors & Metrics

Testing Data: The model was evaluated on a 20% holdout set from the original data, created using a standard train_test_split with random_state=42.
Metrics: The primary evaluation metric was R-squared (R²) to measure the proportion of variance in sales that the model could predict. Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) were also used to measure prediction error in terms of units sold.

Results

Metric	Value
Training R²	0.6324
Testing R²	0.3402
MSE	7811.96
RMSE	88.39
MAE	69.27

Environmental Impact

Hardware Type: Not tracked (likely trained on Google Colab standard CPU instances).
Hours used: Not tracked.
Cloud Provider: Not tracked.
Compute Region: Not tracked.
Carbon Emitted: Not tracked.

Model Card Authors

Hernicksen Satria, Jeremy Wijaya, Lawryan Andrew Darisang (NexusCrew)

Model Card Contact

Please open an issue in the repository for questions or feedback.

jerewy
/

Model_NexusCrew