Method & citations

What the app does, where it comes from, what we added — and what it cannot do.

The published method it implements

Stock Planner implements the No-Code Artificial Intelligence approach to inventory distortion set out by Jauhar, Jani, Kamble, Pratap, Belhadi and Gupta (2024).

Three elements come directly from that paper:

Automatic algorithm selection. Several regression methods are trained and the one that predicts your history best is kept, rather than a method being fixed in advance.
Random Forest Regression as the model that generally wins on transactional sales data.
The Inventory Balance Point — the paper's contribution — expressed as IBP = Predicted Quantity per Transaction × (Total Transactions ÷ Years).

The original study covered 29 products. This implementation has been run on catalogues of several thousand, which is where the extensions below became necessary.

What we added, and why

1. A benchmark against doing nothing clever

Published forecasting studies compare AI against other AI. They rarely ask the prior question: does the model beat simply using each product's own past average? Stock Planner measures this on your data and reports it as a percentage. When the answer is "no better", the app says so and recommends the simpler number. This is standard practice in forecasting — forecast value added — and its absence from the literature is a real gap.

2. Croston's method for irregular demand

A machine-learning regressor trained on "days when something sold" cannot express "sells four times a quarter, roughly thirty units each time". On a real catalogue a large share of products behave exactly that way, and no commercial no-code AI platform offers a method for them.

Each product is classified by average demand interval and demand variability (Syntetos, Boylan & Croston, 2005). Steady sellers are routed to the machine-learning forecast; irregular ones to the Syntetos–Boylan Approximation, a bias-corrected form of Croston's method (Croston, 1972). Products with too little history get no number at all, which is the honest answer.

3. Standard inventory practice around the forecast

ABC classification, Economic Order Quantity, reorder point with safety stock z·σ·√L, inventory turnover, and supplier lead-time variability. These are textbook techniques, not part of the source paper, and are marked as such.

4. A data-fitness check before any number is shown

The method is applied to whatever a business uploads, so no prior validation — the paper's or anyone's — tells a given user what their own numbers are worth. Every dataset is graded at run time on history length, demand density, forecastability and model value added, and the app states plainly what it can and cannot support.

5. Corrections applied to the published formula

Fractional years. Counting distinct calendar years understates annual figures when a file covers part of a year. The true span is used instead.
Planning-cycle classification. Demand-pattern cut-offs assume the time bucket matches your replenishment cycle. Applied to raw daily data they wrongly brand a twice-weekly seller as irregular, so the series is bucketed to your actual review cycle first.
Per-product calibration. Tree ensembles pull extremes toward the average, so the highest volume lines are often under-predicted. Each product shows how far the model sits from its own history, rather than reporting a single catalogue-wide accuracy figure.

What it does not do

Stated here rather than discovered later.

Dependent demand. If your material requirements are derived from a production schedule and a bill of materials, that is an MRP calculation, not a forecast. This tool is for independent demand — retail, distribution, spares, consumables and finished goods.
Stockout censoring. Sales records show what was sold, never what was wanted. Demand lost while out of stock is invisible, which biases forecasts downward on your fastest movers. The app flags products where this is likely; correcting it would require daily stock positions that most exports do not carry.
Promotions and price changes. Not modelled. Promotional uplift may be attributed to seasonality.
Multiple locations. Sales are pooled across sites, so a stockout at one branch can be hidden by stock at another.
Perishability. Economic Order Quantity assumes stock does not spoil. For short shelf-life goods the correct model is a newsvendor formulation, which is not implemented.

How it runs

Everything executes in your browser. The regression trees, random forest, gradient boosting, ridge regression and Croston/SBA routines are implemented directly in JavaScript — there is no server, no API and no external service. The application code contains no network calls of any kind, which is verifiable: the source is unminified and readable in your browser's developer tools.

Analysis history is stored locally so the app can measure how accurate its past forecasts turned out to be against what actually sold. That stays on your device and can be exported as a file you keep.

Verification

The implementation is covered by an automated test suite of 300+ assertions, including Croston and SBA checked against hand-worked examples, the bias-correction factor, demand-classification boundaries, locale-specific date and number parsing, and the Inventory Balance Point calculation.

Available on request for academic review, along with the source.

References

Croston, J. D. (1972). Forecasting and stock control for intermittent demands. Operational Research Quarterly, 23(3), 289–303.

Jauhar, S. K., Jani, S. M., Kamble, S. S., Pratap, S., Belhadi, A., & Gupta, S. (2024). How to use no-code artificial intelligence to predict and minimize the inventory distortions for resilient supply chains. International Journal of Production Research, 62(15), 5510–5534. doi:10.1080/00207543.2023.2166139

Syntetos, A. A., & Boylan, J. E. (2005). The accuracy of intermittent demand estimates. International Journal of Forecasting, 21(2), 303–314.

Syntetos, A. A., Boylan, J. E., & Croston, J. D. (2005). On the categorization of demand patterns. Journal of the Operational Research Society, 56(5), 495–503.

Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.

Citing this tool

If you refer to Stock Planner in academic work:

Quantonique (2026). Stock Planner: browser-based inventory forecasting
implementing no-code AI methods with intermittent-demand extensions.
https://quantonique.com/stock-planner/

Please cite the underlying research above as well — the method is theirs, the implementation and extensions are ours.