Norman Li

Syntetos–Boylan Demand Classification Template

Tue, 25 Aug 2026 00:00:00 +0000

The problem

Not all SKUs deserve the same forecasting treatment. A slow-moving spare part and a high-volume production item behave completely differently, and applying a single forecasting method to both guarantees poor accuracy somewhere.

The classic response is ABC segmentation, but ABC only ranks by value — it says nothing about how often demand occurs or how variable demand sizes are. Two SKUs with identical annual value can have completely different demand patterns and need completely different forecasting and stocking policies.

Demand Planning for Industrial, B2B and Project-Driven Businesses

Thu, 20 Aug 2026 00:00:00 +0000

Industrial and B2B demand is driven by projects, tenders, shutdowns and one-off orders — not continuous consumer buying. A practical framework that combines ABC segmentation, demand-pattern classification, customer intelligence, inventory exposure and procurement commitments to make better decisions under uncertainty.

Ecommerce Sales Analytics: A 5,000-Order Dataset Walkthrough

Tue, 18 Aug 2026 00:00:00 +0000

A quick tour of a 5,000-order ecommerce sales dataset — what’s inside, what the data says about categories, regions, payments and pricing, and what to watch out for before using it.

A Practical Guide to Forecast Error Metrics

Fri, 14 Aug 2026 00:00:00 +0000

MAPE is every planner’s default accuracy number — and it quietly misleads on slow-moving and intermittent demand. A practical comparison of MAPE, MAE, RMSE, WAPE and sMAPE, and when to use each.

Exploratory Data Analysis: The Math and Python for Every Step

Fri, 14 Aug 2026 00:00:00 +0000

A visual EDA guide turned into a reference: for each of the nine steps — overview, missing values, descriptive stats, univariate, bivariate, correlation, outliers, distribution, insights — the exact formula and the exact pandas/seaborn function to run.

Demand Exception Reports: Purpose, Review Process, and Practical Challenges

Thu, 06 Aug 2026 00:00:00 +0000

A Demand Exception Report is an analytical tool used in demand planning and S&OP to identify the products that require human investigation. This guide covers exception categories such as forecast variance, bias, spikes, stockout distortions and launch deviations, plus an eight-step review process and the practical challenges of designing a report that planners actually use.

Store Sales Forecasting with SARIMAX: A Full Walkthrough

Wed, 05 Aug 2026 00:00:00 +0000

A complete SARIMAX forecasting pipeline on the Store Sales dataset — from EDA and stationarity testing through model fitting, residual diagnostics, back-transformation, and forecasting with confidence intervals.

Store Sales SARIMAX Forecast

Wed, 05 Aug 2026 00:00:00 +0000

The problem

Daily sales at a grocery chain follow a strong weekly rhythm, but promotions and holidays throw spikes at it that a naive forecast can’t see coming. Baseline forecasts either ignore the weekly seasonality entirely or get blindsided by holiday demand.

Understanding SARIMAX Outputs: A Diagnostic Guide for Junior Analysts

Wed, 05 Aug 2026 00:00:00 +0000

A step-by-step diagnostic guide for junior analysts on evaluating, interpreting, and refining statsmodels time-series reports.

Excel Solver - Returns-Distribution Fitter

Mon, 27 Jul 2026 00:00:00 +0000

The problem

When assets are issued each week and returned over subsequent weeks—such as containers, pallets, tools, and other returnable equipment—planners need a model that captures the return-lag distribution.

In practice, return timing is often represented by a fixed cycle time based on experience. However, this assumption may not reflect actual return behaviour and can lead to inaccurate return forecasts, inventory imbalances, or unnecessary asset purchases.

ABC-XYZ Analysis: A Practical Guide to Inventory Segmentation for Demand Planning

Sat, 25 Jul 2026 00:00:00 +0000

ABC tells you what matters financially. XYZ tells you what you can actually forecast. Together they form the foundation of every sensible inventory strategy.

Demand Planning KPIs: Learning Notes on Forecast Accuracy Metrics

Wed, 15 Jul 2026 00:00:00 +0000

Eight forecast KPIs that come up constantly in demand planning — with the formula, the business meaning, and the trap hidden inside each one.

Reading Theory of Constraints again

Sat, 11 Jul 2026 00:00:00 +0000

Reading about theory of constraints again. Keep circling back to it whenever a planning process feels stuck — the bottleneck is rarely where the complaints are loudest.

TIL: pandas .explode() for BOM flattening

Wed, 08 Jul 2026 00:00:00 +0000

TIL: pandas.DataFrame.explode() solves the bill-of-materials flattening problem cleanly — one list-column per parent, exploded into one row per component, no manual loop.

Why Your Safety Stock Is Lying to You

Thu, 02 Jul 2026 00:00:00 +0000

Every planner inherits the same textbook formula for safety stock. Almost nobody checks whether their demand actually looks the way the formula assumes it does.

Reorder-Point Simulator

Tue, 30 Jun 2026 00:00:00 +0000

The problem

Planners kept asking “what happens to stockouts if we drop safety stock by 10%” and getting an answer three days later, after a spreadsheet exercise. This simulator answers it in a few seconds.

Approach

A discrete-event simulation over historical daily demand: feed in a demand series, a lead time distribution, and a reorder policy, and it plays the policy forward against real (not assumed-normal) demand.

On forecast accuracy metrics

Thu, 25 Jun 2026 00:00:00 +0000

Every team measures forecast accuracy differently, and almost nobody states which one they mean. MAPE punishes low-volume SKUs unfairly; WMAPE fixes that but hides SKU-level misses. Worth writing up properly at some point.

A Demand Planner's Guide to LLM Evals

Thu, 18 Jun 2026 00:00:00 +0000

Evaluating a prompt change is a forecasting problem in disguise: you need a baseline, a held-out test set, and a way to tell signal from noise.

The Join That Broke Procurement

Sun, 24 May 2026 00:00:00 +0000

A silent one-to-many join quietly tripled our open-PO totals for two months before anyone noticed the number looked wrong.

PO Anomaly Detector

Fri, 10 Apr 2026 00:00:00 +0000

The problem

Duplicate and near-duplicate purchase orders were slipping through — same vendor, same amount, submitted twice within a few days. Manual review only caught them after the second payment cleared.

Approach

A nightly SQL job flags PO pairs within the same vendor where amount and description are within a similarity threshold and the submission dates are close together, surfaced in a Power BI report the procurement team checks each morning.

Vendor Scorecard

Sun, 02 Nov 2025 00:00:00 +0000

The problem

Vendor performance reviews ran on whichever spreadsheet the category manager happened to maintain — inconsistent metrics, no history, no way to compare vendors across categories.

Approach

A single Power BI model combining on-time delivery, quality rejection rate, and price variance against contract, refreshed weekly from the ERP and quality systems.

Bullwhip Effect

Mon, 01 Jan 0001 00:00:00 +0000

Example

Retail demand for a product ticks up 10% one week. The retailer, wanting a buffer, orders 20% more from the distributor. The distributor, seeing that 20% jump and wanting its own buffer, orders 40% more from the manufacturer. By the time the signal reaches raw-material suppliers, a 10% demand change looks like a 60%+ swing.

KL Divergence

Mon, 01 Jan 0001 00:00:00 +0000

KL divergence—Kullback–Leibler divergence—measures how different one probability distribution is from another.

For discrete distributions $P$ and $Q$:

$$D_{\mathrm{KL}}(P\|Q)=\sum_x P(x)\log\left(\frac{P(x)}{Q(x)}\right)$$

It can be interpreted as:

The extra information loss incurred when using $Q$ to approximate the true distribution $P$.

Simple Example

Suppose the true coin distribution is:

LLM Eval Harness

Mon, 01 Jan 0001 00:00:00 +0000

Example

A fixed set of representative inputs, each with a rubric of pass/fail checks rather than a single subjective quality score:

RUBRIC = {
    "mentions_driver": lambda text: "because" in text.lower(),
    "correct_direction": lambda text: expected_direction in text,
    "no_hallucinated_numbers": lambda text: not contains_unverified_number(text),
}

Every candidate prompt runs against the same fixed set before it ships, so a change is judged against a stable baseline instead of a handful of examples that happened to look good.

Pandas & NumPy Time Series Cleaning

Mon, 01 Jan 0001 00:00:00 +0000

A quick-reference cheat sheet of the most commonly used functions, organized by workflow stage. Written for use alongside SARIMAX / Chronos-2 / AutoGluon forecasting scripts.

1. Loading & Initial Inspection

pd.read_csv(path, parse_dates=['date_col'], index_col='date_col')
pd.read_excel(path, sheet_name=0, parse_dates=['date_col'])
df.info()                  # dtypes, nulls, memory
df.describe()               # summary stats
df.head() / df.tail()
df.shape
df.dtypes
df.columns.tolist()

Tip: Always parse dates on load with parse_dates= rather than converting after — avoids silent object-dtype dates that break resampling later.

Reorder Point

Mon, 01 Jan 0001 00:00:00 +0000

Example

A SKU sells 42 units a day on average, with a 9-day supplier lead time and 120 units of safety stock:

reorder_point = (42 × 9) + 120 = 498 units

When on-hand plus on-order inventory drops to 498, a new order is placed — not when the shelf is already empty.