Retail Business Intelligence: What the Evidence Shows

In 2002, a synthesis of 40 out-of-stock studies put the average worldwide retail out-of-stock rate at 8.3%. In 2026, a paper by two of the same researchers opened by noting that out-of-stock levels in grocery and non-grocery consumables “stubbornly remain about 8%”, after two decades of point-of-sale systems, replenishment software, shelf-visibility technology and analytics investment.

That is the authors’ framing rather than a fresh worldwide measurement. If it holds, it is a useful place to start: retailers have had scanner data for decades, so the constraint is unlikely to be a lack of data. Whether that data is accurate, joined up and acted on is another matter, and it is the thread running through this guide.

Most figures here come from US retailers (the NRF returns survey and Walmart’s M5 sales data). The technology-adoption figures come from OECD data on EU retail businesses with 10 or more employees. The Nielsen and Gartner figures come from surveys of marketers and of supply-chain leaders across industries, not only retail.

In short, and as our reading of the evidence rather than a tested rule: if you hold stock, start by measuring how accurate your own inventory records are; if returns are a material cost, report them by SKU and supplier; and treat forecasting as the third step, aimed at aggregate planning rather than the shelf. The sections below set out the evidence behind each step.

Retail business intelligence, for the purposes of this guide, means the reporting and analysis a retailer runs on its own sales, stock, returns and channel data to make buying, replenishment and pricing decisions. This guide is about what the published evidence supports: which retail problems analytics measurably improves, which ones it has not visibly moved, and what the difference between those two lists tells you about where to start.

On this page

• Has shelf availability improved?
• What does better forecasting actually buy you?
• Why do retail decisions get reopened?
• Where is the measurable money in retail data?
• How far along is retail, really?
• Why retail media made reporting harder
• How to scope a retail analytics engagement
• Frequently asked questions

Key Takeaways

• A 2002 synthesis of 40 studies put worldwide out-of-stock at 8.3%; a 2026 paper by two of the same researchers describes the level as still “about 8%” in its framing, without re-measuring it.
• A study of nearly 370,000 inventory records in 37 stores of one retailer found about 65% of them inaccurate. That is not a benchmark for your stores, but a replenishment system cannot order correctly against a record that is wrong.
• In the M5 forecasting competition on 42,840 Walmart sales series, the winning method beat a standard exponential-smoothing benchmark by 22.4% across all levels of the hierarchy, but by only about 3.4% at the product-store level, the grain closest to store replenishment. M5 scored forecast accuracy, not stock or service outcomes.
• Only 7.5% of M5 teams beat that benchmark with the forecasts they submitted for final evaluation, and 12.2% with their best submission during the competition.
• US retailers estimated 2025 returns at $849.9bn of merchandise, 15.8% of total sales and 19.3% of online sales; most of the data behind that sits in systems retailers already run.
• Only 32% of marketers in a Nielsen survey measure media spend across digital and traditional channels together, while 65% expect retail media networks to play a growing role.

Has shelf availability improved?

The out-of-stock figure rests on two sources twenty-four years apart, both involving the same two researchers. Only the first is a measurement.

The 2002 study by Gruen, Corsten and Bharadwaj, announced by the Food Marketing Institute, combined 52 studies, 40 of them on the extent of out-of-stocks, covering 661 outlets, 32 fast-moving categories, 71,000 consumers and 29 countries. The worldwide average was 8.3%, with study highs averaging 12.3% and lows averaging 4.9%, close to an 8.2% benchmark published in 1996. The 2026 paper by Corsten, Gruen and LaBrecque surveyed 55 international grocery and non-grocery consumables retailers. It treats the roughly 8% level as its starting problem rather than re-measuring it worldwide, recommends 95% on-shelf availability as a working target, and concludes that high performers are distinguished not by any particular signalling technology but by how well sensing tools, execution routines and frontline empowerment fit together.

So the two figures differ in scope and method and are not a like-for-like trend. Together they suggest the problem has not gone away, not that it has been measured as unchanged.

One mechanism that can leave shelves empty has been measured well. DeHoratius and Raman examined nearly 370,000 inventory records across 37 stores of one retailer and found about 65% of them inaccurate: the system quantity did not match what was physically there. A replenishment system reading a record that says four units are on the shelf will not reorder, whether or not the shelf is empty.

This is the part worth sitting with before buying anything. Automated replenishment reads the perpetual inventory record directly. Demand forecasting and assortment analysis lean on it less directly: to read sales history correctly they need to know when an item was actually available, and a wrong record can hide stockouts. If your record is wrong as often as it was in that study (about two times in three), better mathematics applied to it produces a more confident wrong answer. The order of work matters more than the sophistication of it.

The published evidence points to operational fixes rather than new technology alone. DeHoratius and Raman found that auditing reduced record inaccuracy while store complexity and distribution structure increased it; the 2026 paper credits disciplined execution routines and empowered store staff. Cycle counting, receiving discipline and exception reporting on the record itself are unglamorous work that analytics supports rather than replaces.

The failure modes below are common ones from general retail practice, not findings of the studies above, and each has a different fix. Stock arrives and is received against the purchase order rather than against what was physically in the carton. An item is moved to a promotional end-cap and is technically in the store but not where the picker or the customer looks. A unit is sold under the wrong barcode, which creates two errors at once: one item overstated, another understated. Shrink removes inventory silently. Returns are credited to the customer before the unit is inspected and put back, so the system believes it is sellable when it is in a bin behind the counter.

Analytics is genuinely useful against this list, but as detection rather than prediction. Most of those failure modes leave a signature in data the retailer already has: negative on-hand quantities, items showing sales after reaching zero stock, SKUs with no movement for a period in a category that turns weekly, receiving variances clustered on one supplier or one shift. Exception reports on those patterns tell an operations manager where to count. They are candidate tests, not proven fixes, so check them against physical counts in one category before rolling them out. Even so, that is a much smaller brief than a demand-forecasting program, and it improves an input that replenishment and forecasting both depend on.

What does better forecasting actually buy you?

This question has an unusually good answer, because retail forecasting has been subjected to a public, blind competition on real data.

The M5 competition used 42,840 hierarchical series built from Walmart unit sales, from the total down to 30,490 individual product-store series. Entries were scored out of sample and compared with benchmarks, the strongest of which was a bottom-up exponential-smoothing method (ES_bu). The results that matter for a buyer:

The uncertainty companion competition adds a sharper point: at the most extreme quantiles (including the upper tail, where demand spikes live), the top 50 methods were on average less accurate than the ARIMA benchmark. Both competitions scored forecasts, not stock outcomes: neither tested reorder policies, service levels or inventory cost.

None of this says forecasting is pointless: the gain is real at aggregate planning levels and small at the shelf. A vendor promising transformative accuracy improvement at SKU-store grain is promising far more than the best M5 teams achieved there: the strongest top-50 product-store score beat the benchmark by about 4.4%.

Ask any forecasting proposal two questions: at what grain is the improvement claimed, and against what benchmark. Ideally there are two baselines: a standard statistical method, which tests the model, and your current process, which shows what would change for you. A gain over your spreadsheet alone does not tell you whether a simple standard method would have done as well.

Why do retail decisions get reopened?

Accuracy is one half of a reporting problem. The other half is how long the organization takes to act on a number once it has it, and as of October 2026 we found much less published evidence on it.

Gartner surveyed 151 supply-chain leaders between November and December 2025, published in July 2026. Respondents had recently made a significant network investment decision and worked at companies with at least $250M revenue across manufacturing, life sciences, retail and technology. 72% had revisited final approval on network decisions at least once, and more than half had revisited three or more times, contributing to delays and lower satisfaction with the outcome. Gartner attributes much of this to day-to-day turbulence (fluctuating demand, labor variability, cost swings), and argues that revisiting a decision is not in itself a failure.

The sample is cross-sector and about network investment decisions. The survey does not measure reporting cycles or test BI as a remedy, so read it as directional at most. Our own view, which the survey does not test either, is that some re-approval loops in retail may have a simpler cause: two functions arrive with different numbers for the same thing.

That is a hypothesis you can check before building anything: ask merchandising and finance to calculate margin for the same scope and cut-off date. If the numbers differ, a dashboard will not fix it. Agree who owns the definition, which cost components are in it, and who settles a disagreement. That conversation is short compared with the build, and it can prevent three teams each building their own margin report.

Where is the measurable money in retail data?

Returns. At an estimated 15.8% of US retail sales, they are one of the largest numbers in retail operations, and in our reading they get less attention in BI proposals than forecasting does.

The NRF and Happy Returns (a UPS company) 2025 report surveyed 2,006 consumers who had returned at least one online purchase and 358 e-commerce professionals at US merchants with more than $500M revenue. Retailers estimated that 15.8% of 2025 sales would be returned, totalling $849.9bn (down from 16.9%, or $890bn, in 2024), with an estimated 19.3% of online sales returned and 9% of returns judged fraudulent. These are survey-based estimates made during 2025, not audited totals, and the co-author sells returns services, but the retailers’ own systems are where the underlying numbers live.

Returns are tractable because most of the data already exists and the decision is close to it: return rate by SKU, supplier, channel, reason code and customer cohort can be built from systems retailers already run, provided orders, refunds, reason codes and restocking can be linked and reason codes are used consistently. Check that before calling returns a quick win. The actions that returns reporting supports (fixing sizing information, delisting a chronically returned line, changing a photograph, identifying return-fraud patterns) are decisions someone can actually take. Rank them by what returns cost, not by return rate alone: a line that is returned often can still pay its way once margin, resale recovery, handling, freight and fraud are counted. Likewise, the $849.9bn is the value of merchandise returned, not the loss.

Compare that with assortment and pricing optimization, a familiar theme in retail analytics marketing, where the published evidence we found is thin. One reasonably rigorous study (a synthetic-control design over 592 treated stores against roughly 2,400 control stores from five other retailers across 104 weeks) estimated a 3.7% store revenue uplift from beer-category assortment recommendations, against 0.3% for a placebo retailer. That is a useful result, but it covers one retailer and one category, reports no confidence interval, says the revenue figures it prints are illustrative rather than actual data, and the authors work for the supplier whose recommendations were tested (AB InBev). As of October 2026 we found no independent replication. Treat it as encouraging rather than established.

How far along is retail, really?

The official figures give a fairly modest picture.

The OECD’s January 2026 analysis of EU retail and wholesale, based on the OECD ICT Access and Usage by Businesses database (enterprises with 10 or more employees), tracks the actual adoption curve:

Among EU retail businesses with 10 or more employees, cloud computing reached 39% by 2023, up from 14% in 2014. Online order-taking reached 43%. AI adoption moved from 7% to 12% between 2020 and 2024, though the OECD notes that more than 30% of Dutch retail businesses reported using AI. Retail is behind wholesale on basic digital presence: 67% have a website against 85% in wholesale. The EU figures are unweighted averages of the countries in the OECD data, so a small member state counts as much as a large one.

None of these indicators measures analytics maturity directly, and as of October 2026 we did not find an official-statistics series that isolates it for retail; the OECD figures are the closest measurement we could use. Read together, they suggest digital capability in EU retail is uneven, so it is safer to check a retailer’s data foundations than to assume an advanced analytics baseline. Anyone quoting a precise “X% of retailers use analytics” figure should be asked where it comes from.

Why retail media made reporting harder

A retail media network makes a retailer an advertising business as well, which adds a reporting surface rather than consolidating one.

Nielsen’s 2025 Annual Marketing Report surveyed 1,400 brand marketers (manager level and above, marketing budgets of at least $1M) online between 25 February and 6 March 2025. Only 32% measure media spend holistically across digital and traditional channels together (23% in Europe), while 65% said retail media networks would play a growing role in their media strategies. Nielsen, itself a measurement company, lists lack of transparency in newer channels such as retail media among the obstacles to cross-media ROI.

These are brand marketers, not retailers, but the gap lands on the retailer’s side of the table: can retail media results be reconciled independently against your own sales data? That is a measurement-design question more than a tooling one, and it is considerably easier to build into the data model early than to retrofit after a year of campaign data has accumulated in a vendor’s reporting portal.

How to scope a retail analytics engagement

For a retailer that holds its own stock and replenishes stores, our reading of the evidence above suggests a sequence that puts forecasting third rather than first. It is a heuristic built from separate studies, not an order anyone has tested head to head.

Other business models may start elsewhere. Three questions decide it: where is the most value at stake, what data does that decision depend on, and can you measure a baseline and test a change against it? Our suggestion, not a tested rule:

A useful test for a forecasting or replenishment proposal is whether it asks about the state of your inventory record. One that opens with forecasting sophistication and never asks how accurate your stock positions are has skipped a well-measured mechanism behind empty shelves. It is not the only one, but it is the one you can check in your own data first. For other use cases, ask the equivalent question about the data they depend on.

If you want outside help choosing that first step, the closest fit among our packages is a 2–4 week current-state and gap assessment with a prioritized roadmap; our retail and e-commerce page covers the sector more broadly. Whoever does the work, insist the first deliverable is a measured number about your own data rather than a roadmap alone.

Frequently asked questions

Is retail BI different from general business intelligence? The tooling is largely the same. What differs is the grain and the data-quality profile: retail decisions happen at SKU-store-day, the perpetual inventory record can be badly wrong (about 65% of records at one retailer in a large study), and if you sell through marketplaces or run retail media, part of the commercially interesting data sits in platforms you do not control. Those three points change sequencing more than they change technology.

What should a retailer build first? If you hold stock and replenish stores, a measured inventory-record accuracy baseline; if returns are a material cost and your order, refund and reason-code data can be linked, returns reporting. Both largely use data you already have and produce a number someone can act on. Other business models should start with the decision that has the most value at stake and a baseline they can measure. Dashboard design matters less at this stage than getting the underlying numbers right.

Does AI-based demand forecasting beat traditional methods? In the M5 competition the winning machine-learning method beat an exponential-smoothing benchmark by 22.4% across all levels combined, but by only about 3.4% at product-store level, the grain closest to store replenishment. M5 measured forecast accuracy, not stock or service outcomes, so ask any vendor at what grain, and against which benchmark, their improvement is measured; the figures are in the forecasting section.

How much should a retail analytics project cost? No universal figure follows from the evidence here. In our view the main drivers are the number of source systems and the state of the data, then history, refresh frequency, security, user groups and rollout support. That is the case for a contained assessment ahead of a build: it lets whoever does the work price the implementation against your actual data rather than an assumption.

Is RFID the answer to inventory accuracy? Not on its own. RFID is a sensing technology, and the 2026 on-shelf availability study found that high-performing retailers are distinguished less by which shelf-visibility technology they adopt than by whether sensing, execution routines and staff empowerment are aligned. Tags still have to be read and acted on, so receiving discipline still matters, and so do counts, prioritized by item value, movement and the error rates you find.

What to measure first

Twenty-four years separate a measured 8.3% average and the round “about 8%” a 2026 paper uses to frame its question. That is not a precise trend line, but it suggests the problem has not gone away while retail added e-commerce data, wider RFID use, cloud warehouses and machine learning to the scanner data it already had.

That is not an argument against retail analytics. It is an argument about sequence: order each use case by the value at stake, the data it depends on and whether you can test it, not by how well it demos. The evidence does not pin empty shelves on one cause: the 2026 paper points to how sensing, execution and staff empowerment fit together, but inventory record error is a well-measured mechanism, and for a retailer that holds stock it feeds replenishment directly and forecasting indirectly.

If you hold stock, start by measuring how wrong your record is, and say how: the share of records that match a physical count exactly, the average unit error and the value-weighted error give different numbers, so state which one you used and any tolerance. A first pass can be as simple as physically counting a sample of SKUs in a few categories and stores and comparing the result with the system record. Keep the sample honest: pick SKUs at random within value and movement bands, count without seeing the system quantity, and use one cut-off time so stock in receiving or in transit is treated consistently. The error rate that comes back, with its margin of error, will tell you a good deal about what to build next.

Related Articles

• BI by Sector: A Complete Guide for Dutch Mid-Market Companies — Business intelligence use cases by sector
• Building a KPI Dashboard: From Data to Insight — From data to a working KPI dashboard, step by step

Sources

Peer-reviewed

• Corsten, Gruen & LaBrecque, Beyond technology: what differentiates on-shelf availability performance in retail, International Journal of Retail & Distribution Management, online-first 28 May 2026. Survey of 55 international grocery and non-grocery consumables retailers.
• DeHoratius & Raman, Inventory Record Inaccuracy: An Empirical Analysis, Management Science 54(4), 2008. Nearly 370,000 records across 37 stores.
• Makridakis, Spiliotis & Assimakopoulos, M5 accuracy competition: Results, findings, and conclusions, International Journal of Forecasting 38(4), 2022. 42,840 Walmart series; 5,507 teams. Level scores for the winner and ES_bu are in Table 2 of the published version; an earlier PDF version lists a different ES_bu score at product-store level (0.926 instead of 0.915).
• Makridakis, Spiliotis, Assimakopoulos, Chen, Gaba, Tsetlin & Winkler, The M5 uncertainty competition: Results, findings and conclusions, International Journal of Forecasting 38(4), 2022.
• Nidhin, Prakash & Safil, Measuring the impact of assortment change in retail stores using the synthetic control method, Journal of Business and Retail Management Research 18(2), April 2024. 592 treated stores, ~2,400 controls, 104 weeks. Authors are employed by AB InBev, the supplier whose assortment recommendations were tested. The paper states that the revenue figures it prints are illustrative and do not represent actual data.

Official statistics and analyst research

• OECD, Local Retail, Global Trends: How Digital, Green and Skills Shifts in the EU are Reshaping SMEs in Towns and Cities, 14 January 2026. Based on the OECD ICT Access and Usage by Businesses database (enterprises with 10+ employees); EU values are unweighted means of the countries shown.
• Gartner, Gartner Survey Shows 72% of Supply Chain Leaders Revisit Final Approvals for Network Decisions at Least Once, Causing Delays, 14 July 2026. 151 leaders surveyed 11 November – 21 December 2025, at companies with at least $250M revenue across manufacturing, life sciences, retail and technology.

Trade bodies and platform data

• Gruen, Corsten & Bharadwaj, Retail Out-of-Stocks: A Worldwide Examination of Extent, Causes and Consumer Responses, Grocery Manufacturers of America / Food Marketing Institute, 2002. 52 studies (40 on out-of-stock extent), 661 outlets, 32 categories, 71,000 consumers, 29 countries. The study was funded by Procter & Gamble. As of October 2026 we could not find an official online copy of the report; its results are reported in Corsten & Gruen, On Shelf Availability: An Examination of the Extent, the Causes, and the Efforts to Address Retail Out-of-Stocks, 2005.
• National Retail Federation and Happy Returns, 2025 Retail Returns Landscape and press release, 15 October 2025. Consumer survey n=2,006; retailer survey n=358. Figures are retailer estimates.
• Nielsen, 2025 Annual Marketing Report, 20 May 2025. 1,400 global marketing professionals, fielded 25 February – 6 March 2025.

Claims we checked and did not use

• “$1.75–1.77 trillion lost to inventory distortion worldwide.” Comes from IHL Group’s commercial inventory-distortion studies ($1.77 trillion in the 2023 edition), updated periodically and widely redistributed by software vendors. The full methodology sits in paid reports, and we could not verify it independently.
• “Returns consume 20–25% of revenue.” Published by a returns-insurance vendor; the metric is ambiguous and does not match the NRF retailer estimate of 15.8% of sales.
• “Around 15,000 inaccurate product-data issues per year” and “10–30% product return rates by sector.” Stated on a standards body’s marketing page without methodology, sample or citation that we could find.
• “60–85% of BI and analytics projects fail.” The versions we traced led to consulting and vendor blogs citing one another; we could not find a named primary study behind them, and the figure varies between 60% and 85% depending on who is quoting it.