Demand Forecasting: Data, Methods, Accuracy & Supply Chain Guide需求预测深度指南：数据、方法、准确率与供应链决策工作流

Quick Answer: What Is Demand Forecasting?快速回答：什么是需求预测？

1. Define Demand Before Building a Forecast1. 构建预测前先定义“需求”

“How much will customers want next month?” is not yet a forecast specification. Demand might mean orders placed, order lines confirmed, units shipped, units consumed, paid subscriptions started, service requests received or lost sales plus fulfilled sales. These measures diverge when supply is constrained, orders are cancelled, customers buy early, shipments are delayed or one order is split across deliveries. Choose the measure that represents the decision you need to support and name the authoritative actuals source.

“客户下月需要多少”还不是完整预测规范。需求可能指下单量、确认订单行、发货量、消费量、付费订阅启动量、服务请求量，或已履约销售加流失销售。当供应受限、订单取消、客户提前采购、发货延迟或一笔订单拆分交付时，这些指标会明显不同。应选择真正支持目标决策的指标，并明确实际值的权威来源。

Write a forecast contract containing the unit of measure, product and location hierarchy, time bucket, horizon, snapshot cutoff, currency or price treatment, returns and cancellation policy, aggregation rule and decision owner. A monthly category forecast for budget planning cannot be evaluated like a daily SKU-store forecast for replenishment. Forecastability usually declines as the horizon grows and the series becomes more granular, so accuracy targets must be tied to a specific horizon and level.

预测契约应包含计量单位、产品与地点层级、时间桶、期限、快照截止点、币种或价格处理、退货与取消政策、汇总规则和决策责任人。用于预算的月度品类预测不能像用于补货的日度 SKU—门店预测那样评估。期限越长、粒度越细，通常越难预测，因此准确率目标必须绑定具体期限和层级。

Demand forecasting is related to, but different from, demand planning. Forecasting produces a statistically or judgmentally supported estimate of unconstrained demand. Planning adds commercial intelligence, consensus review and policy choices, then passes an approved demand plan into supply, inventory and financial processes. The separation matters because a supply constraint should not quietly lower the estimate of what customers wanted.

需求预测与需求计划相关但不同。预测产生由统计或判断支持的无约束需求估计；计划加入商业情报、共识复核与政策选择，再把批准的需求计划传递给供应、库存和财务流程。两者必须分开，因为供应约束不应悄悄降低对客户真实需求的估计。

2. Prepare Demand Data That Reflects Reality2. 准备能够反映真实需求的数据

Historical orders or consumption are the foundation, but raw transactions rarely form a ready time series. Standardize item and location keys, units of measure, timestamps, time zones, cancellations, returns and substitutions. Distinguish order date from requested, promised and shipped dates. Preserve slowly changing attributes such as category, channel and pack size as they were known at each point in time. When product identifiers change, document the mapping rather than silently combining unrelated histories.

历史订单或消费是基础，但原始交易很少能直接形成可用时间序列。需要标准化产品与地点键、计量单位、时间戳、时区、取消、退货和替代关系；区分下单日、要求交付日、承诺日与发货日；保留每个时点的品类、渠道、包装规格等慢变属性。当产品标识变化时，应记录映射，而不是悄悄合并不相关历史。

Observed sales can understate demand when inventory was unavailable, orders were capped or customers substituted another item. Flag stockout periods and estimate lost or deferred demand only with transparent assumptions. Conversely, promotions, panic buying, forward buying and one-time projects may inflate history relative to normal baseline demand. Keep both the original observation and any cleaned or causal-adjusted series so an analyst can reproduce the transformation.

当库存不可用、订单被限量或客户改买替代品时，观察到的销售会低估需求。应标记缺货期间，并只用透明假设估计流失或延迟需求。相反，促销、恐慌性购买、提前采购和一次性项目会让历史高于正常基线。必须同时保留原始观察值与清洗后或因果调整后的序列，使分析人员能够重现转换。

3. Understand Demand Patterns and Forecastability3. 理解需求模式与可预测性

Explore the series before choosing a model. Plot the original level and relevant aggregation, then examine trend, seasonal cycles, calendar effects, structural breaks, outliers, zeros, missing intervals and changes in variance. Separate item launch, growth, maturity, decline and end-of-life phases. A model that works during stable maturity may fail at launch or discontinuation. Check whether apparent seasonality is repeated across enough cycles and whether its timing follows the calendar or shifts with events.

选择模型前先探索序列。绘制原始层级与相关汇总，检查趋势、季节周期、日历效应、结构断点、异常值、零值、缺失区间和方差变化；区分产品上市、增长、成熟、衰退与退市阶段。成熟期有效的模型在上市或退市时可能失效。还要确认季节性是否跨足够周期重复，以及其时间是固定日历还是随事件移动。

Do not apply one model policy to every SKU-location pair. Segment by business value and demand behavior. Smooth, high-volume items may support automated statistical forecasts. Seasonal items need models and horizons that capture recurring cycles. Intermittent items contain many zeros and irregular nonzero demand, so ordinary percentage errors and standard smoothing can mislead. Lumpy, promotion-led or project demand may require scenarios, customer intelligence or aggregation to a more forecastable level.

不要对所有 SKU—地点组合应用同一模型政策。应按业务价值与需求行为分群。平稳高量产品可适合自动统计预测；季节性产品需要捕捉重复周期的模型与期限；间歇性产品含大量零值且非零需求不规则，普通百分比误差和标准平滑法可能误导；块状、促销驱动或项目型需求可能需要情景、客户情报，或汇总到更可预测的层级。

4. A Repeatable Demand Forecasting Process4. 可重复的需求预测流程

1. Frame the decision. Define the target, product-location-time grain, horizon, cutoff and the inventory, capacity, procurement or finance decision the forecast will inform.

   界定决策。定义目标、产品—地点—时间粒度、期限、截止点，以及预测要支持的库存、产能、采购或财务决策。

2. Freeze and reconcile data. Save dated source extracts, filters, master-data versions and transformations. Reconcile the modeled history to authoritative totals.

   冻结并对账数据。保存带日期的来源提取、筛选、主数据版本与转换，并将建模历史与权威总额对账。

3. Diagnose and segment. Identify lifecycle, trend, seasonality, intermittency, promotions, stockouts and structural changes; route series to appropriate model families.

   诊断与分群。识别生命周期、趋势、季节性、间歇性、促销、缺货与结构变化，把序列路由到适合的模型族。

4. Build baselines. Generate naïve, seasonal-naïve or moving-average benchmarks. Advanced methods must beat an honest baseline at the relevant horizon.

   建立基准。生成朴素、季节朴素或移动平均基准；高级方法必须在相关期限上优于诚实基准。

5. Train and backtest candidates. Use rolling forecast origins without future leakage, compare error, bias, interval coverage, stability and maintenance effort.

   训练并回测候选模型。使用无未来信息泄漏的滚动起点，比较误差、偏差、区间覆盖率、稳定性与维护成本。

6. Apply governed adjustments. Add known future events only with documented owner, rationale, amount, duration and evidence; preserve both baseline and adjusted forecast.

   应用受治理调整。只在记录责任人、原因、幅度、持续时间与证据后加入已知未来事件，并保留基准与调整后预测。

7. Publish uncertainty and hand off. Provide point forecasts or quantiles, assumptions and exception flags. Planning systems decide orders, production or allocation under constraints.

   发布不确定性并交接。提供点预测或分位数、假设与异常标记；计划系统在约束下决定订单、生产或分配。

8. Learn from actuals. Join frozen forecasts to observed demand, score by horizon and segment, analyze bias and override value, then update the next cycle.

   从实际值学习。连接冻结预测与观察需求，按期限与细分评分，分析偏差与调整价值，再更新下一周期。

5. Demand Forecasting Methods and Selection Rules5. 需求预测方法与选择规则

Method selection should follow the data and decision, not fashion. Qualitative methods—structured expert judgment, market research, Delphi-style elicitation or comparable-product analogs—are necessary when relevant history is absent. Quantitative methods are appropriate when numerical history exists and parts of the past relationship are expected to persist. Combining forecasts can reduce dependence on one model, but only when each component is independently valid and weights are estimated without leakage.

方法选择应服从数据与决策，而不是追逐潮流。当缺少相关历史时，需要结构化专家判断、市场研究、德尔菲式意见收集或可比产品类比等定性方法；当存在数值历史且部分过去关系可能延续时，定量方法更合适。组合预测可降低对单一模型的依赖，但前提是每个组成模型独立有效，且权重估计不存在信息泄漏。

Evaluate methods at the exact hierarchy, frequency and horizon used operationally. A model may win on monthly category demand but fail for weekly SKU-location replenishment. A global machine-learning model trained across related series may help sparse items, yet must still beat simple local baselines and remain calibrated through assortment, pricing and channel changes. Complexity is justified only when the gain is robust and material to the downstream decision.

必须在实际运营使用的层级、频率与期限上评估方法。某模型可能在月度品类需求上最佳，却无法支持周度 SKU—地点补货。跨相关序列训练的全局机器学习模型可能帮助稀疏产品，但仍须优于简单局部基准，并在产品组合、价格和渠道变化中保持校准。只有当改进稳健且对下游决策有实质价值时，复杂度才合理。

6. Promotions, New Products and Structural Change6. 促销、新品与结构变化

Promotional demand should be decomposed into a baseline and an incremental effect whenever possible. Record promotion type, depth, placement, channel, eligible products, start and end dates, inventory availability and whether demand was shifted from nearby periods or cannibalized another product. Measuring lift against an unsuitable baseline can overstate impact. Future promotion forecasts also require the planned mechanics to be known; a model cannot use a discount or campaign that has not been defined.

促销需求应尽可能拆分为基准与增量效应。记录促销类型、折扣深度、展示位置、渠道、适用产品、起止日期、库存可用性，以及需求是否从相邻期间提前或蚕食其他产品。用不合适基准计算提升会夸大效果。未来促销预测还必须知道计划机制；尚未定义的折扣或活动无法成为模型输入。

New products have little or no direct history. Use analogs chosen by category, price, channel, launch type and expected adoption curve; separate market size assumptions from distribution and availability; and publish scenarios rather than false precision. Transfer-learning or attribute-based models may help when a business has many past launches, but similarity must be validated. Update frequently after launch and distinguish genuine consumer demand from pipeline fill, initial channel loading and temporary scarcity.

新品几乎没有直接历史。应按品类、价格、渠道、上市类型和预期采用曲线选择类比产品，把市场规模假设与铺货、可用性分开，并发布情景而非虚假精确值。当企业拥有大量历史上市案例时，可使用迁移学习或属性模型，但必须验证相似性。上市后应频繁更新，并区分真实消费者需求、渠道铺货、初始装载与暂时稀缺。

Structural breaks also arise from assortment changes, competitor exits, regulation, capacity shifts, route changes and macroeconomic shocks. Do not “clean” these events as outliers when they represent the new environment. Preserve the event label, decide which historical periods remain comparable, shorten training windows where justified and widen forecast intervals until the new regime is better understood.

产品组合变化、竞争者退出、监管、产能变化、路线改变和宏观冲击也会形成结构断点。如果事件代表新环境，就不能把它们作为异常值“清洗”。应保留事件标签，决定哪些历史期间仍可比较，在有依据时缩短训练窗口，并在理解新机制前扩大预测区间。

7. Measure Demand Forecast Accuracy and Bias7. 衡量需求预测准确率与偏差

Accuracy requires a frozen forecast and a matching actual. Store issue date, target period, horizon, hierarchy level, point or quantiles, model version, baseline, adjusted forecast and assumptions. Score only after the actuals adjustment window closes. Never evaluate a historical forecast with a file that was later revised using information unavailable at the time; that creates look-ahead leakage and makes the process appear better than it was.

准确率评估需要冻结预测与匹配实际值。保存发布日期、目标期间、期限、层级、点值或分位数、模型版本、基准、调整后预测与假设，并在实际值调整窗口关闭后评分。不能用后来加入当时不可获得信息的修订文件评估历史预测，否则会产生前视泄漏，让流程看起来比真实更好。

MAE is intuitive at one scale. WAPE can summarize a portfolio but may hide poor service on low-volume or high-value items. MAPE is undefined or unstable when actual demand is zero or very small, which is common at granular levels. Scaled errors can improve comparison across series. Track signed bias separately because over- and under-forecasts cancel in aggregate error. Report results by horizon, value segment, demand behavior, lifecycle and hierarchy—not only one enterprise average.

MAE 在单一尺度上直观；WAPE 可汇总组合，却可能掩盖低量或高价值产品的差表现；细粒度需求常为零或接近零，此时 MAPE 无定义或不稳定；缩放误差可改善跨序列比较。必须单独跟踪有符号偏差，因为高估与低估会在汇总误差中抵消。结果应按期限、价值分群、需求行为、生命周期与层级报告，而不能只给一个企业平均值。

Use rolling-origin backtesting: at each historical issue date, train with information available before that date, forecast the operational horizon, record errors and move the origin forward. Compare candidates on identical folds. For probability forecasts, evaluate quantile or distributional scores and empirical coverage. A range that is too narrow will miss often; a range that is extremely wide may cover actuals but provide little planning value.

使用滚动起点回测：在每个历史发布日期，只用该日之前可获得的信息训练，预测运营期限，记录误差，再把起点向前移动；候选模型使用相同折次比较。对概率预测，应评估分位数或分布评分与实际覆盖率。区间过窄会频繁漏掉实际值，极宽区间虽然可能覆盖，却缺少计划价值。

8. Forecast Hierarchies, Reconciliation and Horizons8. 预测层级、协调与期限

Supply-chain demand exists in hierarchies: SKU to family, store to region, channel to total and week to month. Forecasts created independently at each level usually do not add up. Bottom-up forecasts preserve detail but can accumulate noise. Top-down forecasts are stable at aggregate level but may allocate poorly. Middle-out and statistical reconciliation approaches balance these properties. Whatever method is chosen, publish coherent numbers so totals equal the sum of the approved components.

供应链需求存在多个层级：SKU 到产品族、门店到区域、渠道到总量、周到月。各层独立生成的预测通常无法相加一致。自下而上保留细节，却可能累积噪声；自上而下在汇总层稳定，却可能分配不准；中间层展开与统计协调可平衡这些特性。无论选择哪种方法，都应发布协调后的数字，使总量等于批准组成部分之和。

Forecast horizon should reflect replenishment and capacity lead time. Near-term forecasts may use current orders and recent demand signals; longer horizons rely more on trend, seasonality, lifecycle and scenarios. Evaluate each horizon separately. A model optimized for one-week replenishment can be inappropriate for a twelve-month capacity decision. Freeze multiple horizons from the same issue date so planners can see how uncertainty expands and which assumptions dominate further out.

预测期限应反映补货与产能提前期。近期预测可使用当前订单和最新需求信号；长期预测更多依赖趋势、季节性、生命周期与情景。每个期限必须分开评估。为一周补货优化的模型未必适合十二个月产能决策。应从同一发布日期冻结多个期限，使计划人员看到不确定性如何扩大，以及远期由哪些假设主导。

9. Worked Example: Forecasting a Seasonal Product9. 示例：预测一个季节性产品

Consider a hypothetical beverage sold through three regions. The business needs a twelve-week SKU-region forecast for procurement. Two years of weekly orders show a summer pattern, but four historical stockout weeks suppressed observed sales. The next period also includes a price promotion. The team keeps raw orders, flags stockouts, creates a documented unconstrained-demand adjustment and builds three candidates: seasonal naïve, exponential smoothing with trend and seasonality, and regression using promotion and holiday variables.

假设某饮料在三个区域销售，企业需要用于采购的未来十二周 SKU—区域预测。两年周度订单显示夏季模式，但四个历史缺货周压低了观察销售；下一期间还有价格促销。团队保留原始订单，标记缺货，用有记录方法调整无约束需求，并建立三个候选：季节朴素法、含趋势与季节性的指数平滑，以及使用促销和节假日变量的回归模型。

Rolling backtests show that the regression model has the lowest portfolio WAPE during known promotions, while seasonal naïve is more stable in ordinary weeks and on one sparse region. The team uses a segmented policy rather than declaring one universal winner. It generates a baseline, then adds the approved future promotion effect. The published package includes the point forecast, an 80% interval, stockout-adjustment flag, promotion assumption and reconciliation to the national total.

滚动回测显示，在已知促销期间，回归模型的组合 WAPE 最低；普通周和一个稀疏区域中，季节朴素法更稳定。团队采用分群政策，而不是宣布一个全局赢家。先生成基准，再加入批准的未来促销效应。发布包包含点预测、80% 区间、缺货调整标记、促销假设，以及与全国总量的层级协调。

Procurement does not automatically order the forecast quantity. It combines the forecast distribution with current inventory, open orders, lead-time variation, minimum quantities, shelf life, service targets and supplier capacity. After twelve weeks, the team scores the frozen baseline and adjusted forecast separately. If the promotion override improved one region but harmed two, that evidence changes the next adjustment policy instead of encouraging a broad manual uplift.

采购不会自动按预测数量下单，而是把预测分布与当前库存、在途订单、提前期波动、最小起订量、保质期、服务目标和供应商产能结合。十二周后，团队分别评估冻结基准与调整后预测。如果促销调整改善一个区域却损害两个区域，这一证据应改变下一次调整政策，而不是继续进行宽泛人工上调。

10. AI Demand Forecasting With a Governed Analytics Workflow10. 受治理分析工作流中的 AI 需求预测

AI can support demand forecasting by discovering relevant series, generating transformation and query plans, classifying demand patterns, comparing models, detecting anomalies and explaining forecast changes. These benefits depend on governed inputs and reproducible evaluation. A model should reveal the source tables, cutoff, filters, feature availability, training window, backtest folds and error by segment. A compelling chart is not sufficient evidence if an analyst cannot reproduce the result.

AI 可通过发现相关序列、生成转换与查询计划、分类需求模式、比较模型、检测异常并解释预测变化来支持需求预测。这些价值依赖受治理输入与可复现评估。模型应披露源表、截止点、筛选条件、特征可获得性、训练窗口、回测折次与分群误差。如果分析人员无法重现结果，再漂亮的图表也不是充分证据。

InfiniSynapse should be positioned as an analytical and intelligence layer for this workflow. It can connect governed ERP, order, inventory and warehouse data; support natural-language questions across sources; plan and generate queries; surface intermediate results; compare segments and preserve an evidence trail. It does not replace demand-planning execution, automatically issue purchase orders, schedule factories, allocate stock or operate a supplier portal. Human owners approve definitions, adjustments and operational actions.

在这一工作流中，InfiniSynapse 应定位为分析与智能决策层。它可以连接受治理的 ERP、订单、库存与数仓数据，支持跨源自然语言问题，规划并生成查询，展示中间结果，比较细分并保留证据链。它不会取代需求计划执行、自动下采购单、安排工厂、分配库存或运营供应商门户。定义、调整与运营行动仍由人类责任人批准。

Connect this workflow back to the supply chain analytics guide. Downstream teams can use dedicated guidance on inventory forecasting and inventory optimization. Tool-selection intent should remain on the separate demand forecasting software page rather than turning this guide into a vendor list.

该工作流应回链到供应链分析指南。下游团队可继续阅读库存预测与库存优化专题。工具选型意图应保留在独立的需求预测软件页面，避免把本指南变成供应商清单。

11. Demand Forecasting vs Sales Forecasting11. 需求预测与销售预测的区别

Demand forecasting estimates customer need by product, location and time for supply-chain decisions. Sales forecasting commonly estimates bookings, contracts or revenue from opportunities and sales activity. A sales forecast may be organized by account, seller or territory and can include pipeline probabilities; a demand forecast often includes recurring consumer demand that never appears as a CRM opportunity. Revenue also combines quantity with price, mix, recognition and currency rules.

需求预测按产品、地点与时间估计客户需要，用于供应链决策；销售预测通常根据商机和销售活动估计订单、合同或收入。销售预测可能按客户、销售人员或区域组织，并包含管道概率；需求预测往往还包含从未进入 CRM 商机的重复消费者需求。收入还会叠加数量、价格、组合、确认与汇率规则。

The two should be reconciled through shared product, customer, channel, time and scenario mappings, not forced into one undefined figure. A sales-led launch scenario can inform demand planning; observed sell-through can challenge pipeline optimism; supply constraints can limit fulfilled sales without changing unconstrained demand. Preserve each measure and document the bridge so disagreements become analyzable rather than political.

两者应通过共同的产品、客户、渠道、时间与情景映射进行对账，而不是强行合并为一个定义不清的数字。销售主导的上市情景可输入需求计划，实际动销可检验管道乐观程度，供应约束可限制已履约销售却不改变无约束需求。必须保留各自指标并记录转换桥梁，使分歧可分析，而不是变成政治争论。

12. Common Mistakes and Implementation Checklist12. 常见错误与实施清单

• Define target, grain, horizon, cutoff, actuals source and decision owner.定义目标、粒度、期限、截止点、实际值来源与决策责任人。
• Reconcile orders, units, returns, product-location master data and calendar.对账订单、单位、退货、产品—地点主数据与日历。
• Flag stockouts, promotions, outliers, lifecycle and structural changes.标记缺货、促销、异常值、生命周期与结构变化。
• Segment series and build naïve or seasonal-naïve baselines.对序列分群，并建立朴素或季节朴素基准。
• Backtest candidate methods with rolling origins at operational horizons.在运营期限上用滚动起点回测候选方法。
• Track MAE or scaled error, WAPE, signed bias and interval coverage by segment.按细分跟踪 MAE 或缩放误差、WAPE、有符号偏差与区间覆盖率。
• Preserve baseline, adjusted forecast, override evidence and model version.保留基准、调整后预测、调整证据与模型版本。
• Reconcile hierarchy totals and separate unconstrained demand from supply decisions.协调层级总量，并区分无约束需求与供应决策。

Frequently Asked Questions常见问题

Sources and Evidence Notes资料来源与证据说明

The methodological guidance is grounded in official platform documentation and an expert-maintained forecasting textbook. Product examples are hypothetical; organizations should validate their own data definitions, constraints and governance obligations.

方法说明基于官方平台文档与专家维护的预测教材。产品示例均为假设；组织应验证自身数据定义、约束与治理义务。

• Microsoft Learn — Demand forecasting overviewMicrosoft Learn — 需求预测概述
• AWS — Demand Forecasting whitepaperAWS — 需求预测白皮书
• Forecasting: Principles and Practice — Basic forecasting steps《Forecasting: Principles and Practice》— 预测基本步骤
• Forecasting: Principles and Practice — Time series cross-validation《Forecasting: Principles and Practice》— 时间序列交叉验证
• Forecasting: Principles and Practice — Point forecast accuracy《Forecasting: Principles and Practice》— 点预测准确率