The History of Sales Forecasting

By Victor Hoang · Updated Oct 25, 2026

In April 1960 Peter Winters of the Carnegie Institute of Technology published a paper on forecasting sales. Computers now ran inventory for whole companies. Every product needed a sales forecast, again and again. His method had to be quick and cheap.

Forecasting the money to come is older than selling software, and older than the word "forecast" in a sales meeting. Egyptian officials read the Nile's flood level to estimate the harvest and the tax. A Harvard group sold subscribers a business forecast from 1919. The throughline: each generation found a cheaper way to turn yesterday's records into tomorrow's number, and each one discovered that the number is only as honest as the people feeding it.

Today's forecast call, where a sales leader asks her reps which deals will actually close, mixes three things the history below separates: a gauge that reads a signal, a formula that extends a trend, and a group of people who each hold a private opinion.

Timeline at a Glance

The Flood Gauge Years (c. 250 BCE-1410)

Rulers read a river to price the harvest

The scene is the Delta, in Egypt, in the third century BCE. Where the ancient city of Thmuis stood, archaeologists found stone steps that belonged to a nilometer. National Geographic reported in 2016 that the structure was likely built then and used for about a thousand years to read the river's level. The level worked as a proxy for the strength of the river, and so for the harvest. Seven cubits, roughly 10 feet, was the best height for prosperity. Robert Littman of the University of Hawaii says a strong harvest signal meant "taxes would be higher."

The practice outlasted the pharaohs. A water intake at Rawda Island in Cairo, built in 716 and rebuilt in 861, was used to measure the annual flood level in order to assess the taxes on crops, according to the MIT course material. It was a forecast in the plainest sense: read a measurement today, estimate the yield, set the demand.

A different habit shows up in medieval Italy. Francesco Datini, born in Prato around 1335, returned home in 1382 and founded a holding company of limited companies in Avignon, Florence, Pisa, Genoa, Barcelona, Valencia and Palma de Mallorca. The Prato archive holds a correspondence of about 150,000 letters from 1361 to 1411. A firm spread across seven cities could only be run from reports sent in writing. The same record-keeping thread is traced in the history of CRM.

Be clear about what these are. The Egyptian officials and the Prato merchants were precursors, not sales forecasters. Nobody scored a lead or weighted a pipeline. What they establish is the oldest forecasting pattern: a trusted signal, a record, and someone who acts on the estimate. The problem they left was that the signal was a single number, read by the same people who benefited from it.

For teams today: a forecast starts with a measurement you trust. If the measurement can be nudged by whoever is judged by it, you have a gauge that argues back.

The Barometer Years (1919-1930s)

A Harvard group sells the first business forecast

The turning point is a subscription. According to an NBER working paper by Ray Fair, Matthew Shapiro and Kathryn Dominguez, the Harvard Economic Service offered businessmen a comprehensive forecasting service beginning in 1919. A year's subscription cost $100 and brought a Weekly Letter, a Quarterly Review of Economic Statistics and special supplements. Charles Bullock had founded Harvard's Committee on Economic Research in 1915, the half-academic, half-private home for the work.

The product was built on a chart. The service's first editor, Warren Persons, created the Index of General Business Conditions, whose three curves stood for speculation, business and money. The analysts believed the speculation curve led the business curve, and the paper says they specified a lag: declines in speculation signaled that business would begin to fall in six to twelve months. In sales terms, it was a leading indicator, and the idea survives in every pipeline-coverage ratio. Persons is also credited, on the Economics Website profile, with pioneering methods of seasonal adjustment and trend elimination. He left Harvard in 1928, shortly before the barometer project collapsed.

The failure is the lesson. Fair, Shapiro and Dominguez studied the Harvard forecasts and those of Yale's Irving Fisher for 1929 and the early 1930s and found that both were systematically too optimistic, though they also report that nothing in the data suggests the optimism was unwarranted. Our reading: a forecaster who sells a forecast to subscribers, and who hopes for a recovery, is not neutral about the number.

For teams today: a leading indicator is a hypothesis about the future. Test the lag with your own data, and keep the people who gain from good news away from the dial.

The Smoothing Years (1944-1970s)

A Navy tracking trick becomes a sales forecast

The turning point is a war and a computer. In 1944, according to Everette Gardner's survey of the field, Robert G. Brown was assigned to the antisubmarine effort and asked to build a model for tracking submarines; his tracking model was essentially simple exponential smoothing. During the early 1950s he extended it to discrete data, trends and seasonality, and one early use was forecasting demand for spare parts in Navy inventory systems. Smoothing needed less data storage than moving averages, which is why it spread through Navy inventory systems during the 1950s. Brown presented the work to the Operations Research Society of America in 1956, and it became his 1959 book, Statistical Forecasting for Inventory Control.

Charles Holt, working independently with support from the Office of Naval Research, developed a method for additive trends and another for seasonal data. His memorandum of 1957 went unpublished for decades, according to Gardner, but his ideas reached a wide audience in 1960. In April of that year Peter Winters, at the Carnegie Institute of Technology, published his paper on forecasting sales. His abstract explains the problem: computers had brought mechanized inventory control and production planning, and forecasts "must be made on a routine basis for thousands of products." The method he presented was meant to be quick and cheap, and to compare favorably with more traditional ones. Holt's methods tested on real data became known as the Holt-Winters system.

A second branch grew at RAND. The Delphi method was developed there in the early 1950s to obtain a reliable expert consensus. Norman Dalkey and Olaf Helmer described an experiment in a 1962 RAND memorandum, an abridgement of a report from what RAND called "Project Delphi." A 1963 Management Science paper gives the idea in a line: a group of experts answers a series of questionnaires interspersed with controlled opinion feedback. RAND later used it to forecast trends as far as 50 years ahead. Our reading, not the authors': the modern forecast call, where managers hear each other's numbers and revise, rhymes with Delphi, minus the anonymity.

The third branch is about products without history. Frank Bass of Purdue published a growth model in Management Science in January 1969. Its basic assumption is that the timing of a consumer's first purchase is related to the number of previous buyers, resting on innovative and imitative behavior. Tested on eleven consumer durables, it gave good predictions of the sales peak and its timing, and the paper develops a long-range forecast for color television sets.

What this era solved was scale and discipline: a formula could forecast thousands of items the same way every week. What it left open was the human side. A smoothing formula needs a history, and a new product or a new sales team has none.

For teams today: a simple method applied every week often beats a clever one applied rarely. See sales forecasting methods for how these families compare, and demand forecasting for the product-side descendants of Brown and Winters.

The Spreadsheet and Siebel Years (1979-2010s)

The forecast moves onto the desk and into the CRM

The turning point is a classroom in Aldrich Hall. Dan Bricklin, as the Computing History museum quotes him, came up with the electronic spreadsheet in the spring of 1978 as an MBA student at Harvard Business School. He, Bob Frankston and Daniel Fylstra built VisiCalc. It was demonstrated on May 11, 1979 at the 4th West Coast Computer Faire in San Francisco, and Personal Software released it on October 17, 1979. For a sales manager, a forecast stopped being a memo from the finance department. It became a grid she could change herself. More on that tool is in the history of the spreadsheet.

The next shift put the forecast inside the customer system. Siebel's annual report for 1996 lists "product and revenue forecasting, quote generation" among the capabilities of its enterprise sales software, and names Revenue Forecasting and Product Forecasting among the modules of Siebel Sales Enterprise. The forecast was no longer a separate exercise. It came from the same opportunity records reps updated. The stage-and-probability logic that followed is explained in stage-based pipeline forecasting, and the category labels managers use to sort deals in pipeline forecast categories. The wider rise of that software, from ACT! to Siebel, is traced in the history of sales automation.

Salesforce took the pipeline into a browser. Its prospectus says it was founded in February 1999 and began offering its on-demand CRM service in February 2000. It also describes the problem it was selling against: "the difficulty and relatively high cost of implementing and maintaining enterprise applications." How the record system itself evolved is in the history of CRM.

Then came the uncomfortable result. A ten-year study by John Mentzer and Kenneth Kahn, based on a mail survey of 207 forecasting executives, found greater reliance on quantitative techniques but no improvement in accuracy over the decade. A follow-up by Donna Davis and colleagues, published in 2006, reported decreased familiarity with techniques and decreased forecast accuracy compared with the earlier studies. Better tools had not made better forecasts. The data entered by people, and the incentives around it, set the ceiling. The way quotas and territories pushed those incentives is told in the history of sales quotas and territories.

For teams today: a forecast inside the CRM is only as accurate as the stage definitions and the habits behind them. Measure accuracy openly, as laid out in forecast accuracy.

The Machine Years (2016-present)

Prediction is built into the platform, with caveats

The turning point is the moment vendors put prediction inside the sales platform. On September 19, 2016 Salesforce introduced Einstein, and the announcement names predictive lead scoring among its sales features. Scoring a lead and forecasting a quarter are different tasks, but both ask software to estimate from the record what a person once estimated by feel.

The field's own evidence is more careful than the marketing. The M4 forecasting competition used 100,000 series and compared statistical and machine learning methods. Its authors report that, of the 17 most accurate methods, 12 were combinations of mostly statistical approaches, that the best entry was a hybrid of statistical and machine learning features, and that the six pure machine learning methods performed poorly against the combination benchmark. The lesson is not that machines fail. It is that the winners mixed old and new, the way a sales team mixes a weighted pipeline with rep judgment.

The problem this era inherits is the same one the Harvard barometer met: a model can only extend what the records show. Pipeline data that reps update late, stage definitions that drift, and deals that nobody has challenged will be forecast faithfully. For how agents are being built to read pipelines, see the AI forecasting agent blueprint, and for the fundamentals underneath, forecasting fundamentals. The operating function that now owns the number is covered in the history of revenue operations.

For teams today: before you buy a prediction, audit the record it will learn from. And keep a simple baseline, so you can tell whether the model beats it.

Key People

Key Facts

• The Thmuis nilometer was likely built in the third century BCE and used for about a thousand years.
• The Harvard Economic Service began offering forecasts to businessmen in 1919 at $100 a year.
• Brown was assigned to antisubmarine work in 1944; his tracking model was essentially simple exponential smoothing.
• Winters's paper on forecasting sales appeared in Management Science in April 1960.
• Siebel's 1996 annual report lists revenue and product forecasting in its sales software.
• In M4, 12 of the 17 most accurate methods were combinations of mostly statistical approaches.

The Pattern Behind the History

Our synthesis, not published research. Call it Read, Extend, Combine, Distrust:

• Read. Every era begins with a signal someone trusts: a river level, a business curve, a stage in the pipeline.
• Extend. Each generation finds a cheaper way to carry the signal forward: a chart, a smoothing formula, a spreadsheet, a model.
• Combine. The strongest forecasts blend sources: Delphi's panel of experts, the weighted pipeline plus the rep's call, the hybrid that won M4.
• Distrust. Every era also finds that the forecast is shaped by who produces it, from Harvard's optimism to unchanged accuracy in the CRM age.

What It Teaches Teams Today

• Separate the signal from the opinion. Pipeline data is a gauge. A rep's commit is a judgment. Show both.
• Keep a boring baseline. A smoothed trend line costs nothing, and it tells you what your fancier method must beat.
• Blend methods. A weighted pipeline, history and a manager's call each fail in different ways, so combine them.
• Define stages tightly. Software cannot fix a stage nobody agrees on.
• Score the forecast. Track accuracy by rep and by quarter. Studies of practice have found that more tools alone did not raise it.

For the wider story of how selling reached this point, read the history of sales.