Product Forecasting &
Inventory Optimization

01The Right Problem Statement

The goal is not to minimize forecast error. The goal is to minimize the cost of being wrong.

This distinction, which Vandeput makes forcefully in the opening chapters of both books, is the intellectual foundation of everything that follows. Statistical accuracy — a lower RMSE, a tighter confidence interval — is instrumentally valuable only insofar as it reduces the cost of incorrect inventory decisions. An excellent forecast that informs a bad stocking policy is worse than useless. A mediocre forecast that is correctly translated into inventory decisions can be highly profitable.

For Grow Fragrance, this reframing is not merely academic. When the planning process runs on a founder's growth intuition baked into a rules-based engine, the question to ask is not "how accurate is the engine?" but "what is the cost distribution of its errors, and in which direction does it tend to fail?"

The Newsvendor Problem — The Executive Lens

The Newsvendor model is the single most important concept in this entire document. It is named for a newspaper seller who must decide each morning how many papers to order before knowing how many will sell. Order too few, and you lose potential profit. Order too many, and you're stuck with unsellable inventory.

The model is deceptively simple and extraordinarily powerful. It translates every forecasting and inventory decision into a single, intuitive question: given the cost of ordering too much vs. too little, how many units should I stock?

The critical ratio tells you, directly, what percentile of the demand distribution you should stock to. If CR = 0.80, stock the 80th percentile of expected demand. If demand is normally distributed with mean 500 and std dev 80, stock approximately 567 units (500 + 0.84 × 80).

Cost Asymmetry at Grow

Grow has confirmed that stockouts are materially more painful than overstock. The cost structure for a DTC fragrance brand is characteristically asymmetric:

02Understanding Your Demand Signal

Before choosing a model, understand what your data is actually telling you.

Vandeput dedicates significant attention to demand characterization in Data Science for Supply Chain Forecasting. The choice of forecasting model is downstream of the shape of demand — not upstream. A SARIMA model applied to intermittent, lumpy demand will perform worse than a naïve method.

Demand Decomposition

Any demand time series can be decomposed into four components:

Intermittent & Lumpy Demand

Vandeput distinguishes demand types using two axes: demand frequency (how often does a non-zero demand period occur?) and demand variability (when demand does occur, how variable is the quantity?).

Data Quality Requirements

Vandeput is explicit: no model can overcome bad input data. For Grow, the minimum viable data requirements are:

• Sales history: Daily or weekly sell-through by SKU (not orders — actual sales). Starting from 2024 means roughly 1–2 years of data. Sufficient for seasonal models, limiting for multi-year trend estimation.
• Promotions & events log: Date, SKU affected, type (influencer, paid ad, email blast, sale). Without this, promo demand looks like noise and inflates forecast error.
• Stockout log: When was a SKU out of stock? Censored demand (periods where you would have sold more but couldn't) is one of the most common forecast biases.
• Lead times: Supplier lead time by raw material. This feeds directly into safety stock calculation.

03Forecast Error Metrics

Metrics are not goals — they are instruments. Know what each measures and what it cannot measure.

The Metric Catalog

Translating Forecast Error into Inventory Cost

This is the bridge between statistical modeling and business value — and it is the bridge most implementations never build. Vandeput argues that forecast error is only meaningful when expressed in dollars, not abstract statistical units.

The chart above shows the critical insight: the optimal stocking quantity is not at the mean of the forecast distribution. It shifts right (toward higher stocking) as C_u increases relative to C_o. For Grow, given the confirmed asymmetry, the optimal quantity will consistently sit above the forecast mean — meaning a 50th-percentile point forecast systematically under-stocks.

04Forecasting Models

Start simple. Add complexity only when simplicity provably fails.

Baseline Models — The Floor

Vandeput's most important methodological point on model selection: every model must beat a naïve benchmark to justify its complexity. The naïve forecast is simply: "next period's demand equals this period's demand." It requires no estimation, no parameters, and no data beyond the last observation.

Exponential Smoothing — The Workhorse

Exponential smoothing models are Vandeput's recommended starting point for most supply chain applications. They are computationally lightweight, interpretable, and adapt to changing demand levels over time. The alpha (α) parameter controls how quickly the model responds to new information.

ARIMA / SARIMA

ARIMA (AutoRegressive Integrated Moving Average) models demand as a function of its own past values and past forecast errors. SARIMA extends this with seasonal differencing. These are appropriate when the demand series shows autocorrelation — i.e., knowing yesterday's demand genuinely helps predict today's.

Vandeput is pragmatic about ARIMA: it is more complex to tune (p, d, q parameters plus seasonal equivalents), requires stationarity testing, and rarely outperforms Holt-Winters on typical supply chain data. However, for Grow's DTC channel where day-of-week effects, promotional echoes, and holiday patterns create structured autocorrelation, SARIMA can add value — particularly as data accumulates through 2025 and 2026.

ML-Based Forecasting

Machine learning models (gradient boosting, random forests, neural networks) can incorporate external features — promotions, social media signals, day-of-week, holidays, new product indicators — that classical time series models cannot. Vandeput discusses these in the context of feature engineering for supply chain, noting that their value is primarily in capturing causal demand drivers rather than in time series pattern recognition per se.

Model Selection Framework

The question is never "which model is best in theory?" It is always "which model produces the lowest expected inventory cost on this SKU given this much data?"

05Anchors & Faders — Product Lifecycle Modeling

Every SKU in a fragrance portfolio follows a lifecycle. Modeling the shape of that lifecycle is more powerful than fitting a time series in isolation.

The existing rules-based engine implicitly encodes a lifecycle assumption: new fragrances peak and decay, while a handful of "anchor" products maintain or grow. This intuition is statistically grounded — it maps directly to two well-studied models: the Bass Diffusion Model (for faders) and S-curve / logistic growth (for anchors). Making these curves explicit and fitting them to historical data turns founder intuition into testable, refineable parameters.

Bass Diffusion Model — For New Launch Faders

The Bass model (Frank Bass, 1969) describes the adoption of a new product as the interplay between two populations: innovators who adopt independently of others, and imitators who adopt because of social influence from existing adopters.

For Grow, the critical parameter is m — this is precisely where the founder's growth input enters the model. The Bass curve shape (governed by p and q) can be fitted from historical launch data. Once fitted, all subsequent launches share the same curve shape; only m varies. This is the statistical validation layer for the growth assumption: "given p and q fitted from your previous launches, what value of m is implied by our historical peak demands, and how does that compare to the growth target?"

S-Curve / Logistic Growth — For Anchor SKUs

Anchor SKUs — the 1–2 hero fragrances that grow in the early years before plateauing — follow a logistic growth curve. Unlike Bass, the logistic model has no decay phase: demand rises, inflects, and saturates at a ceiling.

Bayesian Pooling — Borrowing Strength Across SKUs

When a new SKU launches, it has zero history. The naïve approach is to wait for data to accumulate before forecasting. The Bayesian approach is to use related SKUs as informative priors — shrinking the new SKU's forecast toward the cohort mean, and then updating as actuals arrive.

The mathematical machinery is hierarchical Bayesian modeling — at its simplest, a partial pooling model where individual SKU parameters are drawn from a common hyperprior. In practice, a reasonable approximation for Phase 2 is:

This approach is directly applicable to Grow's new format problem: a new size of an existing fragrance can pool from both the cohort-level prior and the parent fragrance's demand history, with higher initial weight on the parent.

06Inventory Optimization

Forecasting tells you what demand might be. Inventory optimization tells you what to do about it.

Safety Stock

Safety stock is buffer inventory held to protect against demand variability and supply uncertainty during the replenishment lead time. Vandeput's key point: safety stock is not waste — it is the cost of uncertainty. Reducing safety stock without reducing uncertainty just increases the probability of stockout.

Reorder Point & Economic Order Quantity

Newsvendor — Full Parameterization for Grow

Returning to the Newsvendor with full parameterization relevant to Grow's cost structure:

Service Levels — Cycle vs. Fill Rate

Vandeput distinguishes two service level definitions that are frequently confused:

07MRP Output → Layered Forecasting System

MRP is a constraint solver, not a demand forecaster. The two must be architecturally separated and connected deliberately.

Material Requirements Planning (MRP) is a production scheduling and materials planning tool. It takes a demand plan as input and calculates what to produce, buy, and when. The demand plan itself is not generated by MRP — it is fed into it. The current state at Grow is that the rules-based Demand Engine produces the demand plan that feeds MRP. The statistical forecasting layer will eventually sit between raw historical data and MRP, either replacing or augmenting the Demand Engine's output.

MRP Data Extraction & What's Usable

System Architecture — Demand Signal Layer

The Feedback Loop

A forecasting system without a feedback loop is not a system — it is a one-time calculation. The operational value comes from continuous updating: actuals flow back in, forecast errors are measured, model parameters are updated, and inventory policies are recalibrated. At Grow's current scale, this loop can run weekly.

# Conceptual weekly update loop
def weekly_forecast_update(sku_id, new_actuals):
    # 1. Ingest new week actuals
    actuals = load_actuals(sku_id) + new_actuals
    
    # 2. Refit model (SES/HW alpha updates automatically; Bass needs periodic refit)
    model = load_model(sku_id)
    model.update(new_actuals)
    
    # 3. Generate new forecast with uncertainty intervals
    forecast, lower, upper = model.predict(horizon=12, confidence=0.80)
    
    # 4. Recalculate safety stock with updated σ
    ss = safety_stock(sigma_lt=model.sigma_lt, z=cr_to_z(cu, co))
    
    # 5. Update reorder point
    rop = model.mean_lt_demand + ss
    
    # 6. Log forecast error metrics for governance
    log_metrics(sku_id, mase=calc_mase(actuals, forecast), bias=calc_bias(actuals, forecast))
    
    return {'forecast': forecast, 'rop': rop, 'ss': ss}

08Growth Input Validation

The founder's growth input is a prior, not a forecast. The statistical system's job is to interrogate that prior with evidence.

The current Demand Engine uses a growth target as a primary driver — accounting for an estimated 60–70% of output. This is not inherently wrong: growth targets are legitimate inputs to planning. The problem is the absence of a validation mechanism. When the growth input is high, the entire production plan is large. If the growth doesn't materialize, that's working capital tied up in inventory. If it does materialize but wasn't planned for, that's stockouts.

The statistical framework provides three things the engine currently lacks:

Shadow Mode Protocol

Shadow mode is the politically and operationally correct way to introduce statistical forecasting alongside a founder's existing system. The statistical model runs in parallel, produces outputs, and accumulates a performance track record — but does not yet influence decisions. This serves several purposes:

• Trust building: The model's track record speaks for itself. When it outperforms the engine, the evidence is in the log — not an argument.
• Gap documentation: Every week, the log shows: Engine said X, statistical model said Y, actuals were Z. Over a season, patterns emerge.
• No political risk: The engine's authority is not challenged during shadow mode. The founder sees this as "we're also running this other thing" — not "we're replacing your system."
• Parameter refinement: The model is being fitted and tuned in shadow mode. By the time it's ready for influence, it's already battle-hardened on real data.

# Shadow mode logging schema
shadow_log = {
    'week': week_id,
    'sku': sku_id,
    'engine_forecast': engine_output,          # Rules-based engine
    'stat_forecast': stat_model_output,          # Statistical model
    'stat_lower_80': lower_bound,
    'stat_upper_80': upper_bound,
    'actual_demand': None,                         # Filled in retrospectively
    'engine_error': None,                          # Filled retrospectively
    'stat_error': None,                            # Filled retrospectively
    'growth_input': engine_growth_assumption,     # Log this explicitly
    'stat_growth_estimate': stat_growth_ci        # Compare to it
}

09MVP & Iterative Implementation

The company needs something in production now. Don't let the statistically correct be the enemy of the operationally useful.

The path from "no statistical models" to "full closed-loop forecasting system" is not a single jump. It is a sequence of phases, each of which delivers standalone business value and de-risks the next phase. The North Star is profitability — which means the first win needs to be visible within 30–60 days, not 12 months.

10Governance & Monitoring

A forecasting system without governance degrades. Metrics drift, models go stale, and nobody notices until it's expensive.

KPIs That Map to Dollars

Model Retraining Cadence

The Long-Term Vision — Monte Carlo Production Planning

The endgame described by the founder — a Monte Carlo engine with inputs from historical and statistical results, creating a "distribution of possible production scenarios" — is achievable within 12–18 months of this implementation. The architecture is:

This architecture transforms the planning conversation from "how many units should we make?" to "here is the distribution of demand outcomes, here is the cost of each tail scenario, and here is the mathematically optimal production plan given your cost structure." The founder's growth input becomes one of several tunable parameters in the simulation — not a monolithic driver, but a lever the team can test and stress-examine.