Patent Application
20120123963
Market analysis (e.g., market share, market size, etc.) is often performed to allow analysts to better understand effectiveness of various operations (e.g., promotional operations, advertising campaigns, pricing campaigns, etc.) of an enterprise (e.g., business, educational organization, government agency), to detect systematic changes in an enterprise, to determine whether particular products are competing effectively with products of competitors, and to make forecasts with respect to existing products or future products.
The present disclosure provides methods, machine readable media, and systems for market forecasting are provided. An example of a method for market forecasting includes modeling market characteristics of market participants for a type of product and deriving variability of an attribute corresponding to a market characteristic coefficient of the type of product for each of the market participants. The method includes resampling from a distribution of the variability of the attribute for each of the market participants and remodeling the market characteristics of the market participants for the type of product using the resampled attribute. The method includes forecasting future market characteristics of the market participants for the type of product according to the remodeled market characteristics.
For example, the market characteristics can include one or more of market share and/or market size. The attribute can be an average that is specific to each market participant. Accordingly, a variability of the attribute can be derived from its historical data. In some examples, the attribute can be trend-adjusted to arrive at an attribute distribution (e.g., a price distribution). The attribute distribution can be resampled (e.g., using bootstrapping) for each market participant to yield a resampled attribute.
The resampled attribute can be added back to the individual trend, as an attribute input to the market model to forecast market characteristics into the future. Using simulation with recursive sampling (e.g., repeated sampling with replacement if the data is from a set of discrete values, or generating random values from a distribution) and runs, market characteristic distributions can be obtained for a future time period. Point forecasts and confidence interval forecasts can be derived from the market characteristic distributions as points of reference.
For example, the attribute can be average sales price equal to $10 for a particular time period (e.g., a month). The average sales price for the previous month is equal to $12. The standard deviation between the average sales price and a de-trended average sales price for all considered time periods is $2. The variability for a particular time period can therefore have a probability distribution, which is dependent on the standard deviation. A random sample can be drawn from the distribution for a time period. For example, a random sample from the distribution of the previous month is $1.2 and a random sample from the distribution of the particular month is $1. The random sample for a month can be added to the average sales price for the month to result in a pseudo-attribute (e.g., a pseudo-price). Thus, the pseudo-price for the previous month is $13.2 and the pseudo-price for the particular month is $11. These pseudo-prices can be fed back into the market model to arrive at different results (e.g., market characteristics). The resampling can be done repeatedly to arrive at multiple pseudo-prices and therefore multiple results from the model. Such implementations can improve the accuracy of the forecast provided by the model while providing a dynamic measurement of the variability of the attributes over time. Those approaches that rely on using averages may not reflect the variability.
Some previous modeling approaches for performing market analysis, such as discrete choice modeling approaches, use the assumption that attributes (such as a price attribute) do not exhibit a trend change. That assumption may not be valid, since in many competitive markets, product prices can continually decline while capabilities of products are continually being enhanced.
In the detailed description of the present disclosure, reference is made to the accompanying drawings that form a part hereof, and in which is shown by way of illustration how examples of the disclosure may be practiced. These examples are described in sufficient detail to enable those of ordinary skill in the art to practice this disclosure, and it is to be understood that other examples may be utilized and that process, electrical, and/or structural changes may be made without departing from the scope of the present disclosure.
The figures herein follow a numbering convention in which the first digit or digits correspond to the drawing figure number and the remaining digits identify an element or component in the drawing. Similar elements or components between different figures may be identified by the use of similar digits. For example, 346 may reference element “46” in
A device 101 can include and/or receive a tangible non-transitory machine readable medium (“MRM”) 103 storing a set of machine readable instructions (e.g., software) for market forecasting. Although, with respect to
The market models 110 can be communicated by the device 101 to a remote location, such as through a network interface 114 of the device 101 and over a network 116 to a client device (e.g., computing device) 118. The market models 110 can be presented for display in a display 120 of the client device 118, or a report (e.g., a chart, graph, table, etc.) can be generated based on the market models 110 for presentation in the display 120. Alternatively, a display device can be directly attached to the device 101 to allow for presentation of the market models 110, or for presentation of reports produced based on the market models 110.
As used herein, a “market model” can refer to a market characteristic model (e.g., a market share model and/or a market size model). One example attribute is a price attribute. Examples of other attributes include product features, product availability locations, product capabilities, product capacities, etc. A “market participant” refers to an entity that can be chosen by a consumer of goods or services. One example of a market participant is a product that consumers can purchase. Thus, in competitive market multiple products may be offered by various competing enterprises (e.g., businesses, educational organizations, government agencies), from which the consumer can make a choice when purchasing. Another example of a market participant is a service that can be offered by an enterprise. A market participant can also refer to the enterprise itself. Thus, in this latter example, a market can include several competing enterprises that sell competing products and/or services (e.g., commercial products and/or services, educational offerings, government services, etc.).
Market models can be used to perform an analysis of market shares (expressed as percentages) of competing participants. Market size models can be used to perform an analysis of market sizes (expressed in terms of amount, such as total revenue, total profit, etc., of each market participant) of competing market participants. A market size model can refer to a model that represents market demand (expressed in terms of units of a good or service, revenue, etc.) for the market participant. Market share and/or market size models can also be used by analysts to perform forecasting for existing products or for future products.
To provide more accurate market models (e.g., market share models and/or market size models), trends of underlying attributes can be considered when building the market share models and/or market size models. A “trend” in an attribute can refer to some tendency of values of the attribute to increase, decrease, or stay constant. A trend of the attribute can change, which means that attribute values can exhibit differing trends in different time intervals. For example, a trend can change over time.
One example of a trend is a change in price over time. With consumer electronic products, for example, price may decline over the life of each of the products. Another consideration of consumer electronic products is that capacities and capabilities of such products tend to increase while prices decrease. Personal computers provide a good example of this changing trend, where the prices of personal computers that are introduced into the market decrease over the life cycles of the computers while processor speeds and memory capacities of the computers increase. Another example of electronic products that exhibit pricing declines with increasing capacities are storage products, such as memory chips, hard disk drives, flash memory devices, etc.
By taking into account the trend (or change in trend) of one or more attributes associated with various choices (representing market participants) that are available in a market allows for the building of more accurate market models. Market share and/or market size models can be built for both individual market participants and the overall market in the environment of attribute(s) exhibiting trend change.
A market model can be based on a conditional logit model. Utility is a measure of happiness and satisfaction gained from a product or service. Therefore, utility is closely related to a product's attributes (e.g., sales price). Despite the quantitative nature of attributes, utility cannot be fully captured and measured, as there are other factors (e.g., random factors) such as personal preferences and tastes affecting the perceived worthiness of the product or service. Accordingly, utility can be measured by two addends: expected utility and random utility. Maximum utility theory suggests that a customer chooses a particular product or service from a choice list based on the perceived maximum utility of the choice. When the random utility is assumed to have the extreme value distribution, the probability of a generic customer choosing a particular brand product, which is the market share for the brand, will have a conditional logit model or multinomial logit model.
Suppose there are K+1 choices (representing market participants) of a particular type of product (e.g., computers, printers, etc.) from which a consumer can choose. One of the choices is defined as the base or the reference choice, and the other K choices are defined as the alternative choices. For each choice, we denote its market share by pk, and model it by pk=fk({right arrow over (x)}), where fk({right arrow over (x)}) is a function based on a vector {right arrow over (x)} that contains explanatory variables.
In one example, product prices are considered the drivers for market share so that the price attributes of the K+1 competing choices are used as explanatory variables. An attribute is considered a driver for a market characteristic if the attribute affects the market characteristic. Therefore, the number of attributes M=K+1 and the list of the explanatory variables {right arrow over (x)}=(itc,x0,x1,x2, . . . , xk), where itc is the intercept term and is equal the value “1” in an example, and x0,x1,x2, . . . , xK are the price attributes of corresponding choices 0,1,2, . . . , K (e.g., x0 is the price attribute for choice 0, x1 is the price attribute for choice 1, and so forth). The intercept term itc measures the combined effect of all of the explanatory or independent variables when they take the value of zero in the response or dependent variable. The order of choices does not matter, and they can be labeled by the corresponding indices for the market share model pk, (k=0,1,2, . . . , K). In some examples, the objective for demand modeling is to find the best fk({right arrow over (x)}) for al the choices.
If other attributes are also key drivers for the product demand, they can also be included in the explanatory variables vector, {right arrow over (x)}. Thus, for example, if there are two attributes that are drivers for product demand among K+1 choices, then the explanatory variables list x can be expressed as (itc,x0,x1, . . . , xK, xK+1, . . . , x2K+1), where xKK+1 to x2K+1 are the variables for the second attribute. More generally, if N attributes (or drivers) are considered for K+1 choices, then the total number of explanatory variables in {right arrow over (x)}, including the intercept term, is N*(K+1)+1.
The market share models are expressed with the following mathematical constructs. K coefficient vectors {right arrow over (βk)}(k=1,2, . . . K) are provided, where each {right arrow over (βk)} coefficient vector corresponds to a respective one of the alternative choices 1 to K. A {right arrow over (βk)} coefficient vector is not defined for the base choice (k=0). Each {right arrow over (βk)} is a vector of 1 +(K+1)=K+2 components, where the first component is for the intercept term, and the other K+1 components are for the K+1 choices. The elements of the {right arrow over (βk)} coefficient vector include {right arrow over (βk)}(itc), βk(0), βk(1), up to {right arrow over (βk)}(K). The coefficient βk(0) represents the relative market share of choice k to choice 0 if all choices set their attribute value (e.g., price) to zero. The coefficient βk(1) represents the net effect of unit price change of choice 1 (while the prices of other choices remain fixed) on the relative share of choice k to choice 0. The other coefficients βk(j), j-1 to K, have similar interpretations.
A model p0 represents the model far the base or reference choice, while model pk represents the model for choice k (k=1 to K). Models p0 and pk are defined by Eq. 1 below:
Eq. 1 provides multilogit demand models. In other examples, other types of market share models can be used. In Eq. 1, {right arrow over (βk)}′ represents the transpose of {right arrow over (βk)}, and {right arrow over (βk)}′{right arrow over (x)} is the inner product of vectors {right arrow over (βk)}′ and {right arrow over (x)}. Each component of a {right arrow over (βk)} vector represents a market share coefficient that is to be multiplied with a corresponding attribute on {right arrow over (x)}. If the attribute considered for building the market share models p0,p1, . . . , pk is the price attribute, then {right arrow over (x)} contains the K+1 prices for the K+1 choices.
From the historical data (e.g., historical data 106 illustrated in
The following describes definitions for market size (demand) models, denoted by Dj(j=0,1, . . . K), which represents the total unit demands for the reference choice and the other K choices, respectively. D=D0+D1+ . . . +DK is the total market size, which can also to be modeled.
For example, if a multinomial logit model framework is used, the following relationship can be derived for j=1,2, . . . , K:
According to Eq. 2, it suffices to model D0, since Eq. 2 allows other models Dj(j=1 to K) to be readily derived once D0 is known. For example, a log-linear model that correlates the unit demand for the reference choice (choice 0) with the available attributes set {right arrow over (x)} is used. This set {right arrow over (x)} can be the price attribute set (e.g., average sales prices, ASPs) of all the pertinent K+1 competing choices, similar to {right arrow over (x)} discussed above for the market share models. The log-linear model for D0 is expressed as:
log(D0)=γitc+γ0x0+ . . . +γKxK={right arrow over (γ)}′{right arrow over (x)}. (Eq. 3)
In Eq. 3, γitc,γ0,γ1, . . . , γK represents the market size (demand) coefficients that are to be multiplied with respective price attributes itc (which is the intercept term equal to one), x0x1, . . . , xK. The market size coefficient γ1 represents the effect on demand for choice 0 in response to unit price change for choice 1 while assuming the prices for other choices remain fixed; market size coefficient γ2 represents the effect on demand for choice 0 in response to unit price change for choice 2 while assuming the prices for other choices remain fixed; and so forth. Equivalently,
D
0=exp(γitc+γ0x0+ . . . +γKxK)=exp({right arrow over (γ)}′{right arrow over (x)}). (Eq. 4)
The model for the total market size D then is:
Note the parameter vector {right arrow over (γ)}=(γitc,γ0,γ1, . . . γK) for the market size model is also of length K+2, the same as each of the {right arrow over (β)}k parameter vectors for the market share models, if the intercept term is included in the market size model. Alternatively, the intercept term can be excluded in the market size model, and in that case, {right arrow over (γ)}=(γ0,γ1, . . . , γK) is of length K+1, instead of the same length as each of {right arrow over (β)}k. In other embodiments, other types of market size model definitions can be used.
The method can include deriving variability of an attribute corresponding to a market characteristic coefficient of the type of product for each of the plurality of market participants at block 224. In some examples, the variability of the attribute can be derived for each of a plurality of time periods. Deriving the variability of the attribute can include de-trending an average attribute time series associated with historical data for each of the plurality of market participants. Deriving the variability of the attribute can include determining a trend of the attribute for each of the plurality of market participants.
Determining the trend can include applying a linear regression to the average attribute time series.
Determining a trend can include using a Holt-Winters algorithm. According to the present disclosure, attributes (e.g., average sales price) can be forecast for a particular product using time series methods such as a Holt-Winters algorithm (e.g., triple exponential smoothing, etc.), an auto regressive integrated moving average (ARIMA), and other time series methods.
For example, a Holt-Winters algorithm can be used to smooth data that shows both a trend and seasonality (e.g., periodicity). An example of a Holt-Winters algorithm is given by:
where γ is the observation, S is the de-trended observation, b is the trend factor, I is the seasonal index, F is the forecast at m periods ahead, t is an index denoting a time period, and α, β, and γ are constants that can be estimated to minimize the mean squared error. The present disclosure is not limited to this example of a Holt-Winters algorithm, as others may be used.
For example, an ARIMA model can be applied to data that sows evidence of non-stationarity (e.g., data that shows a trend), where an initial differencing step can be applied to remove the non-stationarity. As used herein, non-stationarity indicates that a probability distribution changes when shifted in time, which can also lead to a change in a mean and/or variance over time. An example of an ARIMA model is given by:
where Xt is a time series of data including real numbers, t is an integer index, L is the lag operator, αi are the parameters of the autoregressive part of the model, the θi are the parameters of the moving average port and the εt are error terms. The present disclosure is not limited to this example of an ARIMA model, as others may be used.
The method can include resampling from a distribution of the variability of the attribute for each of the plurality of market participants at block 226. Some examples can include resampling (e.g., recursively resampling) from the distribution of the variability of the attribute for each of a plurality of time periods. Resampling can include calculating a difference between the trend of the attribute and the average attribute time series and bootstrapping from the difference between the trend of the attribute and the average attribute time series (e.g., when determining the trend includes applying a linear regression to the average attribute time series).
Bootstrapping can include estimating properties of an estimator (e.g., a variance) by measuring those properties when sampling form an approximating distribution (e.g., an empirical distribution of observed data). Bootstrapping can include randomly resampling from sampled data a plurality of times to obtain alternate versions of an otherwise single statistic (e.g., if the sample were used as a whole). Bootstrapping can allow for an estimate of a distribution of a statistic. The present disclosure is not limited to these examples of bootstrapping.
Resampling can include calculating a standard deviation between the average attribute time series associated with historical data and the trend of the attribute. A mean between the average attribute time series associated with historical data and the trend of the attribute can be calculated. The mean can be added to the standard deviation to create a sum on which a Gaussian distribution can be imposed. Thereafter, resampling can include parametrically resampling from the Gaussian distribution.
Resampled attributes (e.g., as resampled by any of a number of methods) can be added to the trend of the attribute to determine a pseudo-attribute. As described herein, remodeling the market characteristics can include the use of the pseudo-attribute.
The method can include remodeling the market characteristics of the plurality of market participants for the type of product using the resampled attribute at block 228. Some examples can include remodeling the market characteristics for each of a plurality of time periods. In some examples, remodeling the market characteristics of the plurality of market participants for the type of product can include using recursively resampled attributes. In such examples, market characteristic distributions can be derived based on the remodeled market characteristics.
When such data is available, remodeling the market characteristics can include inputting a planned future attribute or planned future attribute range (e.g., price range). For example, a particular market participant may know its planned future attributes or attribute ranges. Likewise, a particular market participant may have access to market intelligence including information about another market participant's planned future attribute or attribute ranges.
The method can include forecasting future market characteristics of the plurality of market participants for the type of product according to the remodeled market characteristics at block 230. Some examples can include forecasting future market characteristics for a plurality of time periods.
For example, a market share 344 can be expressed as a percent of the market. For example, the time periods 346 can include quarters (e.g., 04Q1 can represent the first quarter of 2004).
The chart 340 includes historical data (e.g., from 04Q1-09Q3) for each of the market participants 342 as indicated with the solid lines for each market participant 342. The chart 340 also includes modeled data for each of the market participants 342 as indicated by the dashed line for each market participant 342. For example, the modeled data can correspond to values derived from modeling market characteristics of a plurality of market participants 342 for a type of product (e.g., as described above related to modeling at block 222 with respect to
The chart 450 includes historical data (e.g., from 04Q1-09Q3) for each of the market participants 442 as indicated with the solid lines for each market participant 442. As can be seen from the chart 450, the attribute value 448 for all of the market participants 442 includes a downward trend over the time periods 446. Computing attribute variability without de-trending, in such examples, will not yield accurate results. For example, the variability would be dominated by the effect of the trend and would not reflect the true variability of the attribute.
The lines 456 represent the attribute distribution (e.g., price distribution) between the trend 454 and the historical data 452 (e.g., as described above related to deriving variability at block 224 with respect to
As described herein, the modeled forecast data 562 can be obtained from the remodeled market model according to the resampled attribute. Furthermore, as described herein, the modeled forecast data 562 can include planned future attributes or planned future attribute ranges, when such information is available.
Market forecasting according to the present disclosure can be advantageous over non-systematic ad-hoc approaches in forecasting market characteristics. The use of simulation and statistical resampling methods in addressing variability issues in attributes can provide a more robust and stable forecast for market characteristics.
It is to be understood that the above description has been made in an illustrative fashion, and not a restrictive one. Although specific examples have been illustrated and described herein, other component arrangements, instructions, and/or device logic can be substituted for the specific examples shown.
