Forecasting iPhone 17 Pro Demand: From Statistics to Deep Learning預測 iPhone 17 Pro 需求：從統計方法到深度學習

Every evening, after the Apple Store closes, a quiet but important decision takes place. The inventory system needs to answer one question:

How many iPhone 17 Pro devices should be available tomorrow?

At first glance, this sounds like a straightforward forecasting problem. In reality, it is one of the most valuable decisions in retail operations.

Forecast too low, and customers walk into the store only to discover the model they want is sold out. Some may return another day, but others may buy from another retailer or postpone their purchase entirely. Forecast too high, and expensive inventory sits in the stockroom. Unlike groceries, iPhones don't expire, but every unsold device still represents working capital that could have been invested elsewhere.

The forecast therefore influences far more than tomorrow's sales.

It drives the entire replenishment process.

Tomorrow's Demand
        ↓
Inventory Planning
        ↓
Replenishment Order
        ↓
Store Inventory
        ↓
Customer Purchase

For a single Apple Store, the financial impact may be modest. For hundreds of stores worldwide, forecasting errors quickly become millions of pounds tied up in inventory—or millions in missed revenue. The challenge is obvious.

How do we predict tomorrow's demand?

The answer has changed dramatically over the past several decades. As forecasting problems became larger and more complex, the modelling techniques evolved alongside them. In this article, we'll follow that evolution through one practical example: forecasting daily demand for the iPhone 17 Pro.

Rather than comparing algorithms in isolation, we'll see why each generation of forecasting models emerged, what business problem it solved, and when its additional complexity is actually justified.

A Typical Forecasting Problem

Let's begin with a relatively simple scenario. Suppose you're responsible for inventory planning at one Apple Store in London. Every night, you need to estimate tomorrow's demand for the iPhone 17 Pro. Fortunately, you have historical sales data from the past two years.

After plotting the data, several patterns immediately become apparent.

• Weekends are consistently busier than weekdays.
• Sales increase significantly before Christmas.
• Demand spikes immediately after new product launches.
• Demand gradually stabilises several months after launch.
• Day-to-day variation exists, but the overall behaviour is fairly predictable.

If we visualised the sales history, it might look something like this.

Demand

 ▲
 │                          Christmas
 │                           ▲
 │                           / \
 │             Launch       /   \
 │            ▲           /     \
 │            / \         /       \
 │           /   \_______/         \____
 │
 └────────────────────────────────────────► Time

As a data scientist, your first instinct might be to build a sophisticated AI model. However, experienced forecasters usually start somewhere much simpler. The first question they ask is not

"Which AI model should we use?"

Instead, they ask

"Can the historical demand already explain most of tomorrow's sales?" If the answer is yes, then a statistical forecasting model may be all we need.

Stage 1 — Statistical Forecasting

Classical forecasting begins with a simple assumption — demand is not random. Instead, it is made up of several predictable components.

For our Apple Store, those components might include

• an underlying demand level,
• a gradual trend,
• weekly seasonality,
• annual seasonality.

Rather than learning these relationships from data, statistical models describe them explicitly. One of the most widely used approaches is Holt-Winters Exponential Smoothing.

Rather than modelling the entire time series directly, Holt-Winters assumes that demand can be decomposed into several intuitive components.

Demand = Level + Trend + Seasonality

For our Apple Store, Holt-Winters naturally captures patterns such as

• higher weekend demand,
• increasing sales before Christmas,
• gradual changes in long-term demand.

The model is computationally inexpensive, highly interpretable, and often performs remarkably well when demand follows stable seasonal patterns.

However, Holt-Winters assumes that future behaviour can be explained by these relatively simple components. It cannot easily model more complex temporal relationships hidden within the data.

A second family of statistical models approaches forecasting from a different perspective. Instead of decomposing demand into level, trend and seasonality, ARIMA (AutoRegressive Integrated Moving Average) models how each observation depends on previous observations.

Observation
↓
Remove trend
↓
Model remaining correlations
↓
Forecast
↓
Add trend back

ARIMA makes the time series stationary by differencing — regular differencing to remove trend, seasonal differencing to remove seasonality — then model whatever autocorrelation remains in that stationary series using autoregressive (AR) and moving-average (MA) terms. It's modeling how shocks propagate and decay, not a structural level/trend/season split.

Unlike Holt-Winters, which explicitly models trend and seasonality, ARIMA focuses on the statistical dependency between observations. For example, if demand tends to remain elevated for several days after an Apple product launch, ARIMA can capture this temporal persistence through its autoregressive component. When strong seasonal effects exist—such as increased demand every Christmas—the model can be extended to Seasonal ARIMA (SARIMA) by incorporating seasonal autoregressive and moving-average terms.

Although Holt-Winters and ARIMA take different mathematical approaches, they share the same underlying philosophy.

Historical demand contains most of the information required to forecast future demand.

This assumption is surprisingly effective for many retail forecasting problems.

Statistical models offer several important advantages.

• They require relatively little historical data.
• They train quickly on standard CPUs.
• Their predictions are highly interpretable.
• They are computationally inexpensive to deploy at scale.
• They provide strong baseline performance for products with stable demand patterns.

Even today, many supply chain forecasting systems still rely on Holt-Winters or SARIMA because they are robust, explainable, and often difficult to outperform on relatively stable products.

Stage 2 — Machine Learning

Statistical forecasting assumes that historical demand contains most of the information required to predict the future. Now imagine Apple introduces a nationwide trade-in programme for the iPhone 17 Pro. At the same time, student discounts begin, a new colour is released, and a competitor launches a flagship device.

Tomorrow's demand is now influenced by far more than historical sales alone. Neither Holt-Winters nor ARIMA naturally understands concepts such as

• promotions,
• pricing,
• holidays,
• marketing campaigns,
• competitor activity.

Their entire modelling philosophy assumes that future demand can be inferred primarily from the historical demand series.

We now need models capable of learning from multiple business variables simultaneously. This is where machine learning begins.

Instead of asking

"How can we model the demand curve?"

we now ask

"What factors influence tomorrow's demand?"

This seemingly small change marks the transition from time series modelling to supervised machine learning. Rather than modelling the demand series directly, machine learning predicts demand using a collection of explanatory variables, commonly known as features.

For our Apple Store, we might engineer features such as

Yesterday's Sales
Last Week's Sales
7-Day Moving Average
30-Day Moving Average
Weekend?
Public Holiday?
Promotion Active?
Days Since Launch
Price
Competitor Promotion
Weather
...
↓
Machine Learning Model
↓
Tomorrow's Demand

Unlike statistical models, machine learning is no longer limited to historical demand alone. Every feature represents another piece of information that may help explain customer purchasing behaviour.

The simplest machine learning model is Linear Regression. Instead of assuming demand consists of level, trend and seasonality, regression assumes that tomorrow's demand is a weighted combination of multiple explanatory variables.

For example, the model may learn that running a promotion increases expected daily sales by 35 units, while weekends contribute an additional 20 units on average.

This is a major step beyond ARIMA. Statistical models primarily learn patterns within the demand series. Regression learns relationships between demand and business variables.

However, it comes with one important assumption - the relationships are approximately linear. Unfortunately, retail demand rarely behaves that way.

Tree-based models, such as Random Forest, XGBoost, and LightGBM, go a step further. They automatically discover nonlinear interactions—for example, a promotion might have a much larger impact during the first month after launch than six months later, or weekend demand might only increase significantly when a promotion is active. This flexibility makes tree-based models particularly attractive for retail forecasting because they naturally combine many different business signals without requiring complex mathematical assumptions.

Imagine Apple launches a promotion. Will demand always increase by the same amount?

Probably not.

A promotion during Black Friday may generate far more demand than the exact same promotion in February.

Likewise, a student discount may have little impact immediately after a product launch, but become highly effective six months later. These interactions are difficult to capture using a single linear equation.

Decision trees solve this problem differently. Instead of fitting one mathematical equation, they repeatedly divide the data into increasingly similar groups.

Conceptually, the model asks a sequence of questions.

                    Promotion?
                        │
                   Yes / \ No
                     ▼   ▼

                Weekend?
                     │ Forecast = 45
                Yes / \ No
                  ▼   ▼
      Forecast = 120   Forecast = 80

Rather than estimating coefficients, the model learns a hierarchy of decision rules. For our Apple Store, it might discover patterns such as promotions matter much more on weekends, high prices reduce demand only after launch demand subsides, Black Friday behaves differently from ordinary weekends.

These nonlinear interactions are discovered automatically without requiring us to specify them in advance.

Although a single decision tree is easy to interpret, it is rarely accurate enough for production forecasting.

Modern retail forecasting systems therefore rely on ensemble methods, particularly Gradient Boosting, with algorithms such as XGBoost and LightGBM becoming industry standards.

Instead of building one complex tree, gradient boosting builds many small trees sequentially. Each new tree focuses on correcting the forecasting errors made by the previous trees. The final prediction is simply the sum of all individual trees.

Conceptually, the learning process looks like this.

Tree 1
↓
Residual Errors
↓
Tree 2
↓
Residual Errors
↓
Tree 3
↓
...
↓
Final Forecast

Rather than trying to build one perfect model, gradient boosting gradually improves the forecast by learning from previous mistakes. This simple idea has proven extraordinarily effective.

For tabular business data, algorithms such as XGBoost and LightGBM continue to outperform many deep learning models while remaining relatively fast and interpretable.

Stage 3 — Deep Learning

Machine learning represented a major leap beyond traditional statistical forecasting. It could combine historical demand with promotions, pricing, holidays, marketing campaigns and many other business signals simultaneously.

However, despite their impressive predictive power, these models shared one important limitation — Feature engineering.

Before training could even begin, data scientists first had to decide which historical information might be useful. For example,

lag_1
lag_7
lag_14
lag_28
rolling_mean_7
rolling_std_30
days_since_launch
promotion_flag
holiday_flag
...

These features were not learned by the model. They were manually designed by engineers. As the number of products and stores grew into the thousands, maintaining hundreds of handcrafted features quickly became one of the most time-consuming parts of the forecasting pipeline.

Naturally, researchers began asking a new question.

Can the model learn these temporal representations directly from the raw demand sequence, without requiring manual feature engineering?

Learning from Sequences Instead of Features

Rather than feeding engineered variables into the model, an LSTM receives the historical demand sequence directly.

Historical Demand
↓
LSTM
↓
Tomorrow's Demand

Instead of explicitly telling the model to use a 7-day moving average or a 28-day lag, the network learns which historical observations matter during training.

This process is known as representation learning.

Rather than learning from handcrafted features, the model learns its own internal representation of demand dynamics.

Conceptually, every day's sales update an internal memory.

Day 1
↓
Memory
↓
Day 2
↓
Updated Memory
↓
Day 3
↓
Updated Memory
↓
...
↓
Forecast

Unlike traditional neural networks, which process every observation independently, an LSTM carries information forward through time using a hidden state. This hidden state acts as the model's memory, allowing information from previous days to influence future predictions

Why Ordinary RNNs Weren't Enough Early recurrent neural networks (RNNs) already processed sequential data. Unfortunately, they struggled to remember information over long periods.

As each new observation arrived, information from earlier timesteps gradually weakened during training, a problem known as the vanishing gradient. For demand forecasting, this is a serious limitation.

Imagine today's sales are influenced by

• a product launch three weeks ago,
• a promotion that started last Friday,
• Christmas approaching next month.

A vanilla RNN often struggles to retain all of these signals simultaneously.

LSTM was specifically designed to overcome this problem. The key innovation of an LSTM is its memory cell. Instead of blindly passing information forward, the network continuously decides

• what information should be remembered,
• what information should be forgotten, and
• what information should influence the next prediction.

Instead of manually creating lag features, the network gradually learns which historical information deserves to remain in memory.

At first glance, LSTMs seem strictly better than traditional machine learning. After all, they automatically discover temporal features while eliminating manual feature engineering. In practice, however, they introduce a different challenge.

They require a tremendous amount of data.

Our Apple Store sells only one product.

Even with two years of daily sales, we have roughly

365 × 2
≈ 730 observations

This is a tiny dataset for a neural network containing thousands of trainable parameters. Under these conditions, an XGBoost model with carefully engineered features will often outperform an LSTM. The deep learning model simply does not have enough examples to learn robust temporal representations.

When Deep Learning Finally Wins?

Now imagine Apple no longer forecasts demand for a single product in a single store.

Instead, it forecasts

• hundreds of products,
• across hundreds of stores,
• over many years of sales history.

500 Stores
×
300 Products
×
5 Years

↓
Millions of Historical Observations
↓
One Global LSTM

The problem has fundamentally changed.

Although each individual store may contain only a few years of data, the combined dataset contains millions of observations. The network can now learn patterns that generalise across the entire retail network.

For example,

• weekend purchasing behaviour,
• launch demand curves,
• promotional uplift,
• seasonal shopping patterns,
• inventory replenishment cycles.

Knowledge learned from one Apple Store can improve forecasts for another.

This ability to share statistical strength across many related time series is one of the primary reasons deep learning has become increasingly successful in large-scale retail forecasting.

Choosing the Right Tool

One of the biggest misconceptions in forecasting is that newer models automatically produce better forecasts.

They don't.

Each generation of forecasting models emerged because the business problem became more difficult.

The question is therefore not

"Which model is the most advanced?"

Instead, ask

"How much information do I have, and how complex is the forecasting problem?"

The best forecasting model is the one whose complexity is justified by the business—not the one with the most sophisticated architecture. Good supply chain forecasting has never been about building the smartest AI. It has always been about making better inventory decisions.