Forecasting demand and sales. Forecasting consumer demand

Submitting your good work to the knowledge base is easy. Use the form below

Students, graduate students, young scientists who use the knowledge base in their studies and work will be very grateful to you.

How the demand forecast is formed 

All data on actual sales, conducted and planned marketing campaigns, changes in retail prices and other events must be analyzed. The simplest tool for this is Excel program. Thus, the company will receive statistical forecasts of demand. Next, they are selectively corrected by the analyst and submitted for approval to the relevant departments: sales, purchasing, marketing, etc. The final forecast is approved by the company’s management.  

Formation of demand forecast

Why there are inaccurate demand forecasts

The company's forecasts are inaccurate for four reasons:

variability of demand;
subjectivity of adjustments;
inexperience of analysts;
insufficient software functionality.

Variability of demand

For example, demand for cell phones or clothing is most often impulsive, which means it is uneven in different stores - sales volume depends on how skillfully the product is promoted. Accordingly, when planning, it is necessary to take into account what promotions the store will conduct and adjust the forecast depending on the expected results. Bread is a different matter - this product is in demand in any situation. Therefore, it is enough to calculate the statistical forecast; no adjustments are required. 

Subjectivity of adjustments

Often an expert with extensive experience comes to the company and determines demand “by eye”. However, these kinds of “expert” assessments are often erroneous. For example, the American electronics retailer RadioShack found that in 70% of cases, the forecast adjusted by an expert turned out to be less accurate than a statistical forecast obtained based on the average of weekly sales over the last five weeks (the so-called moving average forecast).  

In addition, different departments that adjust forecasts often act inconsistently or deliberately distort the figures so that they can then shift responsibility to each other. A classic example is the confrontation between sales and purchasing departments. The former claim that low sales are due to the lack of goods in stock, while the latter claim that they do not purchase goods because managers still cannot sell them. Accordingly, the sales department tries to overestimate the forecast, and the purchasing department, on the contrary, tries to underestimate it. This is not in the best possible way affects the accuracy of forecasts. 

Inexperience of analysts

In my experience, analysts or planners make less precise adjustments than category managers who know the product very well. In addition, errors in forecasts often arise for a trivial reason, when the analyst does not adequately master forecasting techniques. As a study by Fildes & Goodwin showed, manual changes in forecast values by small amounts do not lead to a significant increase in forecast accuracy, and in some cases even reduce it. The study involved four companies whose activities are related to supply chains; they manually adjusted up to 75% of their statistical forecasts 1 . 

Inaccuracy of software tools

There are many types of demand. It can be stable seasonal, stable non-seasonal, unstable seasonal, unstable non-seasonal and intermittent. The demand for goods with a short life cycle is highlighted separately. In addition to all this, demand is influenced by many external events: marketing promotions, changes in retail prices, changes in product lines, activity of competitors, etc. Combinations of these events can be either complementary or mutually exclusive. Not all software tools are able to take these nuances into account when generating statistical forecasts, and this results in inaccuracies.

Business automation program Class365 will help you make convenient sales forecasting

Determine the nature of demand. There is no universal algorithm that would guarantee accurate forecasts for all types of goods. But there is important rule– first of all, you need to understand the nature of demand for a product: is it impulsive or natural? Once you determine this, it will be easier for you to choose the right methods for managing sales.  

Use a benchmark demand forecast. The question of how accurate a statistical forecast is arises at the very beginning of the forecasting process. To understand whether forecast adjustments are effective, it is necessary to have a reference forecast for comparison. Then the statistical forecast, for example, for tomorrow is compared with the standard. Such a forecast could be a moving average, for example, over the last seven days. You can choose another standard - the main thing is to be guided by the following principles:  

building a reference forecast should not take much time;
the algorithm should be very simple and suitable for all types of demand. For example, it must equally consistently forecast demand for goods with a sales cycle of one week and for goods with a sales cycle of several years.

The results of further adjustments to the statistical forecast must be compared with the reference one. The introduction of a benchmark forecast will change the process of predicting future demand (see rice. 2). 

Set adequate goals demand forecasting. Here is an example of an incorrectly formulated goal: forecast accuracy should be at least 80% for all products. It would be more correct to formulate the goal as follows: the accuracy of the final forecast should be higher than the accuracy of the reference one. Or this: the accuracy of the final forecast must be higher than the accuracy of the one currently used. 

It would seem that in the latest formulations the goal looks more vague, but it allows us to take into account the specifics of demand for different goods. For example, in the sales department mobile phones Mobistar companies for a long time could not overcome the 30% bar in terms of forecast accuracy due to rare and impulsive sales of goods. The deployment of a time series statistical forecasting system made it possible to increase the accuracy of forecasts to 50%. Further building the process of coordinating the final forecast with experts from other departments helped increase the accuracy of forecasts to 60%. The current level of forecast accuracy is about 70%, and this result was achieved by improving data quality 2 . 

If goals are poorly connected to reality, employees begin to adjust forecasts to the available data. This defeats the purpose of all forecasting work.

Don't copy your competitors' goals. Retail chains often look at competitors when forecasting demand. However, this is not always justified. Goals for forecast accuracy should take into account the size of the company, the specifics of its processes, geographical location, breadth of assortment, etc. That is, regional retail network small or medium-sized companies should not rely on the forecasting accuracy of the federal network from the top 10 list. At the same time, it is useful to borrow information about the organization of business processes and their automation from larger competitors.  

Category management in retail: 3 principles for increasing sales

Focus on products that generate the most profit but have the lowest sales forecast accuracy. If you manage to at least slightly improve the accuracy of the sales forecast for goods that generate the main margin, you will receive a significant financial effect by reducing costs. If you increase, even to 100%, the accuracy of the sales forecast for goods with small revenue, the effect will be much less. If there are two products that are comparable in terms of revenue, it is wiser to focus efforts on increasing the accuracy of the sales forecast for the product that has lower revenue. The fact is that if the accuracy of the forecast is already high, then its further improvement will require incomparably greater efforts. If the initial accuracy of the forecast is low, then it is easier to increase it and, therefore, it is easier to get an increase in profit.  

Control the quality of incoming data. Using a benchmark forecast as a starting point and setting adequate targets does not guarantee accurate final forecasts. It is important to control the quality of incoming data. For example, actual sales information may not reflect the actual picture, since there may have been no sales not due to lack of demand, but, for example, due to a lack of goods in stock. In this case, use the average of sales during periods without shortages. Moreover, you should not try to restore demand perfectly accurately - for forecasting there is no fundamental difference whether real demand yesterday was equal to five or seven units of goods. It is enough to know that the average demand was six units.  

Automate processes that affect demand. Many companies do not collect information about the results of their marketing campaigns and do not evaluate their effectiveness. They are sure that any advertising campaign increases sales, but it doesn't. Likewise, many do not track the history of pricing, etc. It is important to build a process for creating a demand forecast for each product group, and better yet, for each product. And here it is required software. However, when choosing it, pay attention to the possibilities for analyzing the impact of external events on demand, such as holidays, various promotions, changes in retail prices, etc. Refuse to manually adjust the statistical forecast for those product categories for which a specialist’s expert opinion does not provide steady improvement in the accuracy of the final forecast. 

Demand forecasting in action: the Nestle experience 

As an example, I’ll tell you about Nestle’s project to build a demand forecasting process. It was carried out jointly with specialists from SAS. A little information: Nestle produces food products, operates in 469 regions in 86 countries, annual turnover is 90 billion Swiss francs. 

The company attaches particular importance to the formation of demand forecasts to a category of goods called “crazy bulls” - these are products with both high sales volume and variability in demand. The “crazy bulls” include, for example, Nescafe coffee. This product is characterized by steady demand, however, to prevent sales volume from falling, stimulating promotions are constantly carried out. 

Nestle came to the conclusion that using only a statistical forecast, as well as only the expert experience of a planner, does not give the desired results. The management set the task to build step by step process forming a demand forecast to increase its accuracy. They acted as follows: 

1. We generated a reference forecast - it was obtained using the method of averaging sales values. 

2. We created a statistical forecast, then the analyst adjusted the data and passed it on to other departments for consideration. They made adjustments, returned the forecast, and the manager approved it. By the way, the forecast accuracy was calculated using the following formula:   Demand forecast accuracy = 1 – |Forecast – Actual| : Forecast. 

Nestle does not open exact numbers, so let's consider a conditional example. Let's say today is the 22nd. The demand forecast made on the 20th for the 21st is 10 units (cans of Nescafe coffee). Actual sales for the 21st were 8 units. The accuracy of the forecast, according to the formula used by Nestle, will be 80% (1 – |10 – 8|: 10).  

3. To achieve high accuracy of demand forecast, hypotheses were formulated for possible events that could affect demand: holidays, postponement of weekends, structural shifts in sales (for example, caused by the crisis), promotions. Experts assessed the impact of each hypothesis on demand and then compared it with the standard. If this increased the accuracy of the forecasts, the hypothesis was taken into account in the forecasting process.

Let me give you a conditional example (unfortunately, Nestle specialists did not provide precise data on what exactly they did as part of the expert adjustment process). The company learned that a competitor unexpectedly reduced prices by 1%. The expert’s experience shows that such actions will lead to a 3% drop in sales. This means that it is necessary to reduce the forecast value by these 3%. 

Let's return to the experience of Nestle. The initially generated statistical demand forecast for the “mad bulls” showed an accuracy of 55.2%. It was then subjected to an expert adjustment process, which increased the accuracy of the final forecast to 82.4%. In addition, the company has increased the accuracy of forecasts for other product categories. All this freed up time for marketers and planners. They began to pay more attention to consistently profitable products, concentrating their efforts on complex products (which require constant promotions, etc., to maintain high demand). Products with low sales are processed on a residual basis.

Formation of demand forecast

1.	Data	Information about sales, balances, deliveries, other movements, as well as marketing promotions and other external events.
2.	Tool	A program that can be used to generate a statistical forecast of demand (in our case Excel)
3.	Statistical forecast	Forecast generated using a demand forecasting tool
4.	Adjusted forecast	Manual adjustment of the statistical forecast by an analyst or planner
5.	Consensus forecast	Manual adjustment and coordination of the final forecast between departments (sales, marketing, etc.)
6.	Approved forecast	Approval of the final forecast by the responsible manager and transfer to departments for implementation.

Segment	Accuracy of statistical demand forecast, %	Forecast accuracy after adjustment by experts, %
“Horses” - products with high sales and low variability	92,1	92,7
“Hares” - products with low sales and high variability	56,3	55,5
"Mad Bulls" - products with high sales and high variability, such as Nescafe brand coffee	55,2	82,4
“Mules” - products with low sales and low variation	90,9	91,2

Each company has its own examples, Nestle discloses information only for the “mad bulls” segment – the Nescafe drink. This table gives directors a reason to think and try to create a similar one for their assortment. After all, Nescafe in Pyaterochka may be in a completely different category than Nescafe in Azbuka Vkusa.

Copying material without permission is permitted if there is a dofollow link to this page

A special feature of measuring demand is the fact that it can be carried out indirectly through indicators of the sale of goods or the population’s expenditures on the purchase of certain consumer goods.

Methods for measuring potential demand

Potential demand - this is the maximum possible amount of demand that potential consumers of a particular product can present. The measurement quantities are:

Number of potential consumers;

Potentially possible number of sales in physical units;

The size of potential sales in value terms.

Potential demand is a theoretically calculated value, which in reality, as a rule, is not achieved. However, its measurement is necessary to compare potential opportunities with actual demand.

There are two approaches to measuring potential demand :

The “search” method consists of intermediate estimates of the calculation of final consumers, based on the total number of people.

The “construction” method consists of first identifying all possible groups of potential consumers and then summing them up.

Methods for measuring real demand

Real demand represents the amount of actual sales of goods for a certain period, expressed in physical or value terms. Since the amount of sales does not fully correspond to the amount of demand, but serves only as an indirect measurement of it, various estimation methods are used.

Calculation based on production volumes, exports, imports and inventories.

Measurements of receipts, sales and inventories of goods in the assortment for a sample of stores over a certain period of time (calculation of the so-called Nielsen index).

Estimates of consumer spending based on budget statistics.

Demand forecasting methods:

Demand forecasting is carried out using various methods. In practice, as a rule, an integrated approach is implemented, taking into account the strengths and weaknesses of the methods used.

There are different methods of demand forecasting general and special appointments.

General methods Demand forecasting is based on extrapolation, expert assessments, standards and economic and mathematical modeling.

1) Extrapolation methods. The simplest methods, based on statistical analysis of time series, make it possible to predict the growth rate of sales of goods in the near future, based on trends that have developed in the past period of time.

2) Methods of expert assessments. The methods are based on obtaining objective assessments as a result of the subjective opinions of experts and specialists. Knowledge of certain procedures is assumed (Delphic method, brainstorming, devils' advocate, etc.).

3) Regulatory methods. They are often used when forecasting demand for industrial goods. The size of purchases is determined by the requirements of technological, construction and similar norms and regulations.

4) Methods of economic and mathematical modeling. The most complex methods that require special training. Based on taking into account the correlation of demand and the factors determining its development.

Special methods forecasting takes into account the characteristics of demand for various goods. They proceed from separate forecasting of the main components of demand and their subsequent summation.

Price Elasticity of Demand shows by what percentage the quantity demanded will change if the price changes by 1%. The price elasticity of demand is influenced by the following factors:

The presence of competing goods or substitute goods (the more there are, the greater the opportunity to find a replacement for the more expensive product, that is, the higher the elasticity);

A change in the price level that is invisible to the buyer;

Conservatism of buyers in tastes;

Time factor (the more time the consumer has to choose a product and think about it, the higher the elasticity);

The share of the product in consumer expenses (the greater the share of the price of the product in consumer expenses, the higher the elasticity).

Products with elastic demand by price:

Luxury items (jewelry, delicacies)

Products whose cost is significant for the family budget (furniture, household appliances)

Easily replaceable goods (meat, fruits)

Products with inelastic demand by price:

Essential items (medicines, shoes, electricity)

Products whose cost is insignificant for the family budget (pencils, toothbrushes)

Hard-to-replace goods (bread, light bulbs, gasoline)

Income Elasticity of Demand shows by what percentage the quantity demanded will change if income changes by 1%. It depends on the following factors:

The importance of the product for the family budget.

Whether the product is a luxury item or a necessity item.

Conservatism in tastes.

Cross elasticity of demand is the ratio of the percentage change in demand for one good to the percentage change in the price of some other good. A positive value means that these goods are interchangeable (substitutes), a negative value shows that they are complementary (complements).

There are fundamental principles, which must be taken into account when forecasting demand at any level of the hierarchy of planning decisions. Failure to comply with these principles leads to the fact that the demand forecast becomes either of low quality or irrelevant from the standpoint of decisions made by the enterprise.

Demand forecasting horizon. The time difference between the moment when demand is forecast and the planning period for which demand is forecast is called a time lag. The choice of the required time lag depends on how much time the enterprise needs to develop and take all necessary measures in order to respond to information about the demand forecast. If it takes a year to increase production capacity in accordance with the forecast of future demand growth, a demand forecast with a horizon of less than a year is insufficient and will not solve the necessary problem of managing production capacity. Or, for example, if the duration of the production cycle for “production to order” is one month, it is illogical to have a shorter forecast horizon, because the enterprise will not be able to respond to such a forecast in time by preparing the necessary reserves of raw materials.

When choosing a demand forecasting horizon, it is necessary to take into account that for more distant planning periods the forecast will be less accurate than for closer ones. Accordingly, the choice of demand forecasting horizon must be justified by those decisions that are made on the basis of the generated forecast - too short a forecasting horizon does not allow adequately solving the problem, and a longer one creates problems with the quality of the forecast.

Selecting a demand forecasting object . The more detailed the forecast, the less accurate it is. Accordingly, for each level of the hierarchy of plans, it is necessary to select the level of detail of the forecast object that will allow solving the task, but does not lead to unnecessary detail. Detail is considered unnecessary, which, while increasing the labor intensity and cost of the demand forecast, does not add value to the forecast from a decision-making point of view.

In general, we can say that the parameters of demand forecasts are determined by the purpose of using the forecast. The higher the level of decision-making and the larger the scale of the decisions made, the larger and for the longer term the demand forecast is built.

Quality of demand forecast. Any forecast has an inherent risk of error. It is difficult to imagine a forecast that does not contain error. Two types of demand forecast error can be distinguished: error in estimating the volume of demand and error in estimating the structure of demand. These types of errors must be considered depending on what level of decision-making in the enterprise we are talking about.

Risk of error in estimating demand volume when forecasting demand can be at any planning level. When forecasting long-term demand, risk manifests itself at the level of product categories and product groups. Risk affects the availability of the required volume of key resources and execution financial plans enterprises. When quickly forecasting demand, the risk manifests itself at the level of product line items and affects the actual level of customer service.

Risk of error in estimating demand structure when forecasting demand, it appears only with long-term forecasting of demand at the level of product categories and product groups. One demand structure is assumed within a product group by item, but in fact the structure turns out to be different.

These risks can be taken into account in two ways: by improving the quality of forecasts and/or by reserving resources specifically intended to cover these risks. In practice, as a rule, they use both methods simultaneously - they work to improve the quality of demand forecasts, and (since one hundred percent forecast quality is unattainable in practice) they form reserve quantities of resources (reserve inventories, reserve time, reserve production capacity).

To assess the quality of a forecast, there are two main measured characteristics: forecast error and forecast accuracy.

Forecast error— the difference between the actual value of demand and its forecast value. It can be expressed both in absolute terms and in relative terms - as a percentage of the actual value of demand.

Forecast accuracy is a percentage value equal to the difference between 100% and the demand forecast error expressed as a percentage.

The main thing for assessing the accuracy and quality of the demand forecast is the measured error and accuracy of the forecast for each individual planning period.

However, more often the interest is not in a single planning period, but in the extent to which a particular demand forecasting method is good. To do this, it is customary to calculate summary characteristics of demand forecast accuracy. The two main ways to evaluate the accuracy of a demand forecasting method are MAPE (Mean Absolute Percentage Erro) and Mean Percentage Error (MPE).

Demand forecast errors can be classified into two categories: random variation and bias.

Random deviations mean that the total forecast errors tend to zero, and planning periods for which demand was overestimated alternate with planning periods for which demand was underestimated, that is, there is no bias in the demand forecast error, negative and positive values Demand forecast errors generally cancel each other out.

Bias means that there is a serious problem - much more serious than random errors - systematic overestimation or underestimation of demand forecast. The forecast bias can be explained by both objective and subjective circumstances. Objective ones include the choice of a demand forecasting model, which may not be entirely relevant, for example, it may not take into account significant factors influencing demand. Objective circumstances can be assessed and adjusted by improving the forecasting model, collecting and preparing data to forecast demand, and training employees in forecasting.

Subjective circumstances are associated with a deliberate underestimation or overestimation of the forecast value. This means that the forecaster is interested, for one reason or another, in shifting the forecast, since he receives certain benefits from the shift in the forecast. For example, if the demand forecast is formed by the sales department of an enterprise, and at the same time it receives a bonus for exceeding the sales plan, it is difficult to expect an optimistic demand forecast from it. Conversely, if the marketing department generates a demand forecast, and the marketing budget is calculated as a percentage of planned revenue, you should not expect a pessimistic demand forecast. Neutralization of the influence of subjective factors is achieved to some extent proper organization demand forecasting process.

In conclusion, it should be said that, in addition to the concept “ forecast accuracy “, we can highlight the concept of “forecast quality”. The quality of the forecast is understood as the ability of the demand forecasting process to generate forecasts that consistently differ from the actual demand values by no more than a given error value. That is, forecast quality means the ability to keep the forecast error within specified limits. This is very important from a management point of view, since the enterprise can prepare for the given limits of the forecast error in advance, and such a scale of error does not jeopardize the level of customer service.

It has been noted that the quality of demand forecast is determined to a greater extent by the good organization of the demand forecasting process than by individual, no matter how complex, mathematical models. However, let's look further at what types of demand forecasting methods exist and in what circumstances it is advisable to use them.

At first glance, the question sounds absurd, but if you take a closer look, you can reveal the following: “if a product has a significant number of facts of zero sales (demand for the product is rare), then all point forecasting methods (including complex ones) will give bad result"

A way out of the situation may be the use of special mathematical modeling methods that allow one to calculate the cumulative probability of demand occurrence. That is, estimate without trying to guess the number of goods sold, but to see with what probability a particular volume of goods can be sold. This will allow us to understand how much product needs to be stored in order to provide a particular level of service.

When simplified, the mechanism is as follows. Special software conducts a series of experiments (100,000 times) about the possible demand for a product during the delivery period (in Western practice - lead time LT). It analyzes how many times demand of different volumes occurred. After this, the cumulative probability of demand distribution is constructed (no more than what volume of goods will be sold with different probabilities)

After this, the service level is taken into account and the optimal inventory is calculated as the demand value corresponding to the cumulative probability equal to the service level.

This can be seen more clearly in the following graph or table: (from Forecast NOW!):

Rice. 1 Level of service and optimal inventory using the example of the Forecast NOW program!

Probability Amount,%	Volume, units

In the figure, the cumulative probability is plotted in light blue. The optimal stock is at the intersection of the set service level and the cumulative probability.

Thus, the use of such methods will help to immediately calculate the optimal inventory for rare goods.

An important issue remains the criterion for classifying goods as rare:

To do this, the average distance in days between adjacent sales facts is calculated. If this number is more than 1.25 days, then we have rare demand; if it is less, it is smooth.

Product sales history:

Average distance between adjacent sales facts = ((3-1)+(4-3)+(7-4)+(8-7))/4 = 1.75 >1.25 -> rare demand

But for goods with smooth demand, demand forecasting is indispensable:

Why do you need to forecast demand?

The work of any trading enterprise is inevitably associated with the problem of optimizing inventory. An excess of goods leads to additional financial costs, and a shortage leads to the loss of regular customers and a decrease in sales volumes. In both cases, there is a shortfall in possible profit, which in conditions of intense competition can cause bankruptcy of the enterprise.

One of the most important components of maintaining an optimal assortment of goods is operational and long-term forecasting of demand. Of course, when planning purchases, you can not predict anything, using the established or emerging level of demand as a source of initial information. However, such an outdated approach in a dynamically changing market and a “spoiled” buyer can hardly be called effective (with the exception of small settlements where there is only one store).

Demand forecasting allows not only to develop an optimal procurement plan, but also to effectively manage enterprise resources. So, for example, knowing that next month there will be increased demand for something other than ice cream, it will be possible to hire salespeople in advance, purchase refrigeration equipment and provide additional financing. If all such events begin to be held during the peak season, then all efforts may be in vain and even unprofitable.

How to forecast demand

To forecast demand, a huge number of theories and special tools have been developed.

Special software

For example, when planning purchases for a supermarket, you cannot do without specialized software. The main problem here is the huge assortment of goods, which is simply physically impossible to “keep in your head.” In addition, special software allows you to automate the process of preparing applications, which, with large volumes of purchases, allows you to save a lot of time.

Microsoft Excel

With a small range of products, excellent results in demand forecasting can be obtained using a standard application Microsoft Excel. Special statistical functions, such as TREND and GROWTH, allow you to instantly process large amounts of information without entering complex formulas. The rich design capabilities of Microsoft Excel will help you present the data obtained not only in tabular form, but also in a more visual form - in the form of graphs and diagrams.

Manually

The demand forecast for individual items of goods can also be compiled manually. So, for example, if a product is new, then even the most sophisticated statistical formulas and previously accumulated information will not help predict its popularity. In such cases, you have to rely not on calculations, but on intuition and additional factors (customer opinions, advertising support, etc.).

Formulas and methods for forecasting demand

The methods used in forecasting demand are very diverse - from naive ones (it is assumed that demand next month will be the same as last month) to the use of complex economic and mathematical theories and their software implementations (neural networks) in calculations.

Simple Average Method

The simplest of these methods is the use of calculations using the “simple average” formula. Demand forecast for next period with this method it is calculated as the arithmetic average of demand indicators for all previous periods. The disadvantage of this method is its high “conservativeness” - outdated information about previous sales will prevent the latest trends demand.

Moving average method

The “moving average” method reacts more quickly to changes in demand. In this case, the calculation is made not on the basis of data for the entire observation period, but for the last several periods.

The key issue is the determination of the “sliding window” - over how many recent periods it is necessary to carry out averaging. The longer this period, the more the moving average forecast coincides with the simple average.

The period can be determined empirically based on the forecast error (RMSE) - calculate this error for different periods and choose the optimal one.

Obviously, the optimal period is 4 days.

An interesting variation of the method is the calculation of the moving average by certain days(that is, for all Mondays, the moving average for the last n Mondays is considered, etc.) This method may be suitable for goods that have a pronounced weekly seasonality (for example, bread).

Weighted average method

A combination of the above methods is the “weighted moving average method”. This model calculates a weighted average based on multiple periods, but gives less weight to more distant periods. Thus, longer observations can be taken for calculations, but the influence of irrelevant data on the calculations can be limited.

Exponential smoothing method

Unfortunately, the above “average” calculation methods allow one to obtain only very approximate results. A more accurate forecast can be achieved by using the “exponential smoothing” and “exponential smoothing with trend” models. In the first method, the latest sales forecast is adjusted based on the forecast error made in last period. The second calculation method (also called the “double exponential smoothing” method) takes into account data with trends - thanks to this, this method can be used even for medium-term forecasting.

Holt-Winters method

To take into account seasonality and the general demand trend, the Holt-Winters model (three-parameter exponential smoothing) is used. To obtain a demand forecast in this method, it is necessary to correctly select three parameters. To do this, you can use either special algorithms or limit yourself to simple brute force.

Autoregression method

If you want to get even more advanced forecasts, you can use "autoregressive" models. This technique allows for a very detailed analysis of the available data, identifying any trends and eliminating random influences. However, unlike previous methods, selecting many parameters will require a lot of effort and time from the user.

Neural networks, genetic algorithms

It should be noted that the more complex forecasting methods are used, the more difficult it is to practical application and the higher the likelihood of errors occurring. Analysis of huge volumes of information, selection of optimal parameters, prompt accounting of market fluctuations - all this is sometimes at the limit of human capabilities. The most promising solution to this problem is the use of “neural network” algorithms. In this technique, a special program, after preliminary training, is able to independently find best solution- at the same time, the user does not need to delve into all the intricacies of the theories used. In addition, “neural networks” are able to take into account hidden trends and create a reliable forecast in such an unstable situation, where previously forecasting was considered completely impossible.

According to research conducted by specialists of the Forecast NOW project, forecasting with neural networks gives better results than all the above methods:

The X axis shows the number of products in the analysis, the Y axis shows what percentage neural networks are better than another algorithm in relative terms.

Rice. 2 Neural networks + Genetic algorithms (GA) and exponential smoothing

Rice. 3 Neural networks + Genetic algorithms (GA) and autoregression

Rice. 4 Neural networks + Genetic algorithms (GA) and Holt Winters method

It can be seen from the figures that forecasting by neural networks gives a significantly better result.

Conclusions

To forecast demand you need:

Determine the nature of demand for a product (if it is smooth, forecasting is needed, if it is rare, forecasting is not needed, you can calculate the optimal inventory using mathematical modeling methods)
Determine methods for forecasting demand (if the product range is small, then you can manually or using Excel, if it is large, it is better to use special software
Determine demand forecasting methods (standard methods work well for some products (see moving average), in general best results achieved by neural networks
It is important to remember that demand forecasting is only the first link in the supply chain, and even the most accurate demand forecast, if inventory and replenishment are not properly managed, will not ensure the efficiency of the entire supply chain.

Submitting your good work to the knowledge base is easy. Use the form below

Similar documents

How the demand forecast is formed

Why there are inaccurate demand forecasts

Demand forecasting in action: the Nestle experience

Why do you need to forecast demand?

How to forecast demand

Special software

Microsoft Excel

Manually

Formulas and methods for forecasting demand

Simple Average Method

Moving average method

Weighted average method

Exponential smoothing method

Holt-Winters method

Autoregression method

Neural networks, genetic algorithms

How the demand forecast is formed 

Demand forecasting in action: the Nestle experience 