If you prefer to read this report in a PDF format, please choose the ‘Print this page’ button on the left side of the page and select the option to print to PDF....
If you prefer to read this report in a PDF format, please choose the ‘Print this page’ button on the left side of the page and select the option to print to PDF.
1. Forecasting data on intellectual property rights ( IPRs ) is important for planning future resource requirements and revenues at IP offices, and for use as part of corporate reporting.
2. The UK Intellectual Property Office ( UKIPO ) applies a “bottom-up” forecasting approach, disaggregating IPR filings by applicant group (country of origin, representation, and technology type) to identify trends. Future trend scenarios by applicant group are selected through discussion with examination groups in the relevant business areas. Performance against forecasts is monitored through business intelligence reporting.
3. This paper identifies forecasting methods used by IP offices and other organisations based on literature available in the public domain, and information requested directly from the IP offices by the IPO. This exercise is undertaken to investigate additional options for future forecasting at the IPO.
4. The following methods are found to have been used or tested by the IP offices and other organisations reviewed:
trend extrapolation has been used or tested by several IP offices (European Patent Office ( EPO ), United States Patent and Trademark Office ( USPTO ), Spanish Patent and Trademark Office ( SPTO )) to forecast future patent applications based on their historic growth, typically using an autoregressive ( AR ) model
), United States Patent and Trademark Office ( ), Spanish Patent and Trademark Office ( )) to forecast future patent applications based on their historic growth, typically using an autoregressive ( ) model Hingley and Nicolas (2004) find AR models perform poorly under certain market conditions, such as financial crises or periods of rapid technological change, when IP filings fluctuate greatly from past trends. Reliance on this technique may therefore pose a risk for IP offices
ii. Improvements to the autoregressive ( AR ) model
using SPTO data, Hidaglo and Gabaly (2012) find that the autoregressive integrated moving average ( ARIMA ) model improves predictive accuracy compared to the simple AR model , by forecasting based on lagged moving averages to smooth out the influence of outliers
data, Hidaglo and Gabaly (2012) find that the autoregressive integrated moving average ( ) model improves predictive accuracy compared to the simple model , by forecasting based on lagged moving averages to smooth out the influence of outliers the EUIPO finds Vector Autoregression ( VAR ) improves upon ARIMA when forecasting EU trade mark ( EUTM ) and registered community designs ( RCD ). In VAR , a dynamic relationship between IP and economic time series is modelled, where all variables are jointly determined. VAR forecasting is also used by financial fa including the European Central Bank ( ECB ) and International Monetary Fund ( IMF )
) improves upon when forecasting EU trade mark ( ) and registered community designs ( ). In , a dynamic relationship between IP and economic time series is modelled, where all variables are jointly determined. forecasting is also used by financial fa including the European Central Bank ( ) and International Monetary Fund ( ) the Organisation for Economic Co-operation and Development ( OECD ) seeks to improve upon the simple AR model by including non-linearities such as discontinuities and structural breaks, relaxing the assumption of continuation of past trends , when producing its macroeconomic forecasts
) seeks to improve upon the simple model by including non-linearities such as discontinuities and structural breaks, relaxing the assumption of continuation of past trends , when producing its macroeconomic forecasts the EPO relaxes the assumption of homogenous trends across applicant countries, and forecasts weights for regional blocs when forecasting subsequent EPO filings. A similar approach could be used to account for industry differences in filing trends
some IP offices use economic theory to select explanatory variables to include in their forecasting models. The EPO uses R&D expenditure to forecast patent filings , based on a production function for knowledge used in contemporary R&D-based endogenous growth models
uses R&D expenditure to forecast patent filings , based on a production function for knowledge used in contemporary R&D-based endogenous growth models using SPTO data, Hidalgo and Gabaly (2013) find controlling for GDP and the industrial production index improves predictive accuracy when forecasting patents and trade marks
data, Hidalgo and Gabaly (2013) find controlling for GDP and the industrial production index improves predictive accuracy when forecasting patents and trade marks EUIPO (2023) find that controlling for industrial sector confidence improves EUTM forecasts and controlling for consumer confidence improves RCD forecasts Consumption and investment from National Accounts improves accuracy of both forecasts
forecasts and controlling for consumer confidence improves forecasts Consumption and investment from National Accounts improves accuracy of both forecasts on Swiss Federal Institute of Intellectual Property ( IPI ) data, Bock et al (2004) find controlling for the Dow Jones index and Swiss consumer confidence index in structural state-space models improves forecasting performance
) data, Bock et al (2004) find controlling for the Dow Jones index and Swiss consumer confidence index in structural state-space models improves forecasting performance on EPO data, Hingley and Park (2016) find that controlling for the cyclical component of GDP can improve predictive accuracy of forecasting patent filings in the presence of shocks
data, Hingley and Park (2016) find that controlling for the cyclical component of GDP can improve predictive accuracy of forecasting patent filings in the presence of shocks WIPO find no improvement in forecasts of PCT (patent cooperation treaty) filings when controlling for GDP, which may reflect that alternative filing routes hold a more complex relationship with changing economic conditions
The EPO and World Intellectual Property Organisation ( WIPO ), as supranational patent offices, use domestic (‘first’) filings as an indicator of subsequent EPO and PCT filingsusing a two-stage transfer function model. This method may be less suited to national offices that attract fewer subsequent patent filings
the EPO and other organisations (including OECD ) accompany forecasting models with annual survey data. Dannegger and Hingley (2013) find the EPO ’s annual survey improves predictive accuracy for a forecasting horizon of one year before declining noticeably . The EPO stopped their large-scale annual survey in 2021 but continue a smaller-scale pulse survey
and other organisations (including ) accompany forecasting models with annual survey data. Dannegger and Hingley (2013) find the ’s annual survey improves predictive accuracy for a forecasting horizon of one year before declining noticeably . The stopped their large-scale annual survey in 2021 but continue a smaller-scale pulse survey potential non-response bias introduced by low response rates to surveys should be considered. The OECD reduces reliance on survey data in its macroeconomic forecasts by producing forecasts as a combination of “soft” indicators, such as consumer surveys, and “hard” indicators, including quantitative empirical data (industrial production, retail sales, house prices etc)
use of expert panels is common across IP offices and other organisations that forecast. Experts intervene by interpreting survey data ( USPTO , EPO ), creating scenarios based on consideration of driving factors (EUIPO, EPO ), and refining model forecasts to account for hard-to-predict events such as legislation changes ( EPO , Euromonitor)
, ), creating scenarios based on consideration of driving factors (EUIPO, ), and refining model forecasts to account for hard-to-predict events such as legislation changes ( , Euromonitor) until 2007, the USPTO used the Delphi method to interpret survey data, whereby an expert panel adjusts their forecasts after each survey round, based on interpretation of the group survey response
viiii. Artificial intelligence ( AI )
Havermans et al (2017) find use of artificial intelligence ( AI ) forecasting techniques by the ETMDN (European Trade Mark and Design Network) improves forecasting accuracy compared to traditional methods. Instead of manual trial-and-error, algorithms select explanatory variables, parameters, lags and optimal transformations to best fit the data, and remove outliers. A new framework based on AI forecasting techniques has been implemented by 22 European IP offices under ETMDN
find use of artificial intelligence ( ) forecasting techniques by the (European Trade Mark and Design Network) improves forecasting accuracy compared to traditional methods. Instead of manual trial-and-error, algorithms select explanatory variables, parameters, lags and optimal transformations to best fit the data, and remove outliers. A new framework based on forecasting techniques has been implemented by 22 European IP offices under AI forecasting techniques are increasingly and successfully used in scientific, economic and business fields (such as energy, market research, big data, engineering, finance, business, biology, health, defence and robotics). Bayesian dynamic modelling has been used by private companies to forecast product sales. Support vector regression modelling (SVR) has been used by energy companies to forecast energy demand. Expectation maximisation ( EM ) has been used by Eurostat to select factors for forecasting unemployment, GDP and inflation. Application of AI methods to IP forecasting is still a novel approach (Havermans et al, 2017)
ix. Error correction model ( ECM )
an ECM estimates the speed at which a dependent variable returns to equilibrium (its long run trend) after a change in another variable, allowing modelling of response to shocks. The Department for Transport ( DfT ) uses an ECM to estimate elasticities of passenger demand for air travel with respect to price, income and GDP
estimates the speed at which a dependent variable returns to equilibrium (its long run trend) after a change in another variable, allowing modelling of response to shocks. The Department for Transport ( ) uses an to estimate elasticities of passenger demand for air travel with respect to price, income and GDP an ECM can only be used if a long run relationship exists between the dependent and independent variables (‘cointegration’). Josheski et al (2011) find evidence of cointegration between quarterly growth of patents and quarterly GDP growth, which may support use of an ECM . EUIPO (2023) conduct a Johansen Cointegration Test that rejects the presence of cointegration between trade mark and design filings and economic variables (confidence indicators, consumption and investment), evidence against use of VECM (vector error correction model)
ix. Online access to IP office forecasting tools
some IP offices have made their forecasting tools accessible and interactive online, to improve transparency. The USPTO ’s Patent Pendency Model is accessible via an interactive spreadsheet on their website, and the trade mark and design forecasting tool developed by the TMDN (the European Trade Mark and Design Network), used by 22 European IP offices, is also available online
The current focus of the IPO’s forecasting effort is on capturing different behaviours within IPR filing data by disaggregating this into different applicant groups. This disaggregation is different depending on the intellectual property right being considered. Examples of disaggregation groupings include country of origin, representation, and technology type.
This approach allows a more meaningful exploration of changing trends than top-level analysis, and the creation of forecasts unique to the disaggregated groups of applicants. It also allows shocks to input among applicant groups to be identified, investigated, and reviewed to determine how best to incorporate into current understanding and future forecasts.
Using this disaggregated data, the forecast process extrapolates trends for these applicant groups and uses historic growth/shrinkage rates to produce a range of possible scenarios. These scenarios are discussed with the business areas responsible for delivering the examination process. This allows discussion and sharing of insights, which are reviewed alongside forecast options, to select the most appropriate option for each disaggregated group.
Business intelligence reporting is used to review performance against forecast and track input for the disaggregated groups. This allows regular monitoring of the input received by the IPO, to better understand the movements and their drivers, and act as a ‘warning system’ should input notably stray from forecast.
When creating input forecasts for patents, these are created for applications, searches, and exams to identify resource requirements and revenue streams associated with these different actions. Similarly, the IPO creates forecasts of renewal income for all rights to predict the monies received, considering likely renewals due, dropout rates, and when payments are made.
The IPO is keen to learn from others by investigating the use of alternative forecasting approaches, subject to being able to find suitable explanatory variables or relevant international data. Where appropriate, these alternative approaches will be incorporated as additional options for future forecasts.
3. Approaches to forecasting: IP offices and organisations
IP Office / organisation Forecasting methods used Forecasting method(s) tested European Patent Office ( EPO ) * Scenario planning
* Autoregressive integrated moving average ( ARIMA ) model
* Annual survey * Econometric modelling European Union Intellectual Property Office (EUIPO) * Historical trend projection
* Vector Autoregressive model ( VAR )
* Impulse response function ( IRF )
* Scenario planning * ITF (Intelligent Transfer Function) Spanish Patent and Trademark Office ( SPTO ) * N/A (information could not be sourced) * ARIMA
* Simple econometric model with a predictive lag variable
* Polynomial Distributed Lag ( PDL ) model
* Intelligent Transfer Function ( ITF ) model Federal Intellectual Property Institute of Switzerland ( IPI ) * Simple trend model
* ARIMA * State-space model with explanatory variables United States Patent and Trademark Office ( USPTO ) * Microsoft Excel simulation tool (Patent Pendency Model)
* Econometric models with explanatory variables
* Exponential smoothing models World Intellectual Property Organisation ( WIPO ) * Trend analysis
* Transfer model * Using economic indicators as explanatory variables The European Trade Mark and Design Network ( ETMDN ) / Cooperation Fund project * Support Vector Machines ( SVM )
* Artificial Neural Networks ( ANN )
to inform its annual budget, a set of scenarios for annual growth of patent applications are presented to the EPO ’s management committee in February each year. By this time, previous-year figures have typically stabilised and can be input to the forecast model. A scenario is selected by the management committee following internal discussion of drivers, and this is then presented to governing bodies between May and June, to inform the end of year budget, approved in December
’s management committee in February each year. By this time, previous-year figures have typically stabilised and can be input to the forecast model. A scenario is selected by the management committee following internal discussion of drivers, and this is then presented to governing bodies between May and June, to inform the end of year budget, approved in December the lag between forecasting and budget approval can cause problems if perception of future market conditions change during the year (Hingley and Nicolas, 2004) , though the forecast is rarely adjusted. In addition to its annual forecast, the EPO forecasts over a 5-year horizon to inform financial planning.
, though the forecast is rarely adjusted. In addition to its annual forecast, the forecasts over a 5-year horizon to inform financial planning. scenario-based planning has been used by the EPO since 2019, prior to which a single forecast of patent application growth would be presented to the president of the EPO for sign-off. The practice of discussing scenarios is now preferred by the EPO ’s management committee, as it allows consideration of current drivers
since 2019, prior to which a single forecast of patent application growth would be presented to the president of the for sign-off. The practice of discussing scenarios is now preferred by the ’s management committee, as it allows consideration of current drivers the forecast model used to produce growth scenarios at the EPO is a two-stage domestic patent transfer model. As a supranational patent office, the EPO mainly attracts subsequent filings, meaning applicants typically first file a patent application at their national patent office . These ‘first’ filings are used to forecast subsequent EPO filings. In the first stage of the transfer model, first filings in Europe, Japan and the USA are forecast using an autoregressive distributed lag (ADL) model based on economic growth theory and the knowledge production function. Transfer ratios of first filings into EPO patent applications are then forecast per bloc of origin using trend extrapolation (see Figure 1). In the second stage, subsequent EPO filings are forecast in an ADL model. This bases forecasts on domestic filings, previous filings at the EPO , size of the EPO market, and economic activity in Europe . The transfer model is solved dynamically using an iterative method (1000 repetitions of the Gauss-Seidel Method) . -in the past, the EPO has used the autoregressive integrated moving average ( ARIMA ) model to inform its patent application growth forecasts. This forecasts filings based on their historical movement, but unlike the simple autoregressive ( AR ) model, uses past moving averages of variables to smooth out the influence of outliers. This model is now only used by the EPO for cross-checking outputs of the transfer function model, and typically does not inform budget forecasts, as it is found to perform poorly in times of turbulence
is a two-stage domestic patent transfer model. As a supranational patent office, the mainly attracts subsequent filings, meaning applicants typically first file a patent application at their national patent office . These ‘first’ filings are used to forecast subsequent filings. In the first stage of the transfer model, first filings in Europe, Japan and the USA are forecast using an autoregressive distributed lag (ADL) model based on economic growth theory and the knowledge production function. Transfer ratios of first filings into patent applications are then forecast per bloc of origin using trend extrapolation (see Figure 1). In the second stage, subsequent filings are forecast in an ADL model. This bases forecasts on domestic filings, previous filings at the , size of the market, and economic activity in Europe . The transfer model is solved dynamically using an iterative method (1000 repetitions of the Gauss-Seidel Method) . -in the past, the has used the autoregressive integrated moving average ( ) model to inform its patent application growth forecasts. This forecasts filings based on their historical movement, but unlike the simple autoregressive ( ) model, uses past moving averages of variables to smooth out the influence of outliers. This model is now only used by the for cross-checking outputs of the transfer function model, and typically does not inform budget forecasts, as it is found to perform poorly in times of turbulence until 2021, the EPO carried out an annual survey on applicants’ filing intentions, that had been run since 1996. 1,000 of the most active applicants, and 4,000 randomly selected applicants were invited to participate, the random group selected such that applicants per bloc and filing type had equal probability of being selected . The survey had a low response rate (35%) , and respondents’ filing intentions were found to serve as a poor indicator of future filings. As a result, the survey was stopped in 2021
carried out an annual survey on applicants’ filing intentions, that had been run since 1996. 1,000 of the most active applicants, and 4,000 randomly selected applicants were invited to participate, the random group selected such that applicants per bloc and filing type had equal probability of being selected . The survey had a low response rate (35%) , and respondents’ filing intentions were found to serve as a poor indicator of future filings. As a result, the survey was stopped in 2021 the EPO continues to carry out a yearly smaller-scale pulse survey on patent filing intentions, most recently in Q4 2022 as part of its user satisfaction survey. Random sampling is conducted within four strata: technology field, user type (applicants and external representatives), country, and filing power (number of applications filed). 234 responses were received by telephone and online to the most recent survey (from 131 applicants and 103 external patent attorneys) Findings are presented to the EPO management committee alongside outputs of the transfer function model, when drawing up forecasting scenarios
continues to carry out a yearly smaller-scale pulse survey on patent filing intentions, most recently in Q4 2022 as part of its user satisfaction survey. Random sampling is conducted within four strata: technology field, user type (applicants and external representatives), country, and filing power (number of applications filed). 234 responses were received by telephone and online to the most recent survey (from 131 applicants and 103 external patent attorneys) Findings are presented to the management committee alongside outputs of the transfer function model, when drawing up forecasting scenarios further methods have been tested for use by the EPO . Hingley and Park (2016) tested a structural econometric model to forecast EPO patent filings amid cyclical shocks. The model took a dynamic log-linear functional form, with components included to control for GDP and R&D spend by technological field. Filings were found to respond more to GDP fluctuations than to deviations in R&D spending from trend. Similar findings were made by Hingley and Park (2017), testing a dynamic model of patenting separating the cyclical component of GDP from its trend component, and finding that patent filings are strongly pro-cyclical. These results suggest forecasting accuracy of models may be improved by controlling for the cyclical component of GDP. Analysis has been commissioned by the EPO to produce a 20-year forecast for patent applications using GDP as the main indicator. This is expected to be published next year
the model specification used to forecast domestic filings in the first stage of the EPO ’s transfer model is informed by endogenous growth theory. Technological progress is assumed to be the driver of GDP growth , and is assumed to evolve according to a knowledge production function whereby research labour input is proxied by R&D expenditure. However, when tested empirically, Hingley and Park (2016) find that EPO patent filings are not sensitive to short run movements in R&D , which may reduce performance of the model
’s transfer model is informed by endogenous growth theory. Technological progress is assumed to be the driver of GDP growth , and is assumed to evolve according to a knowledge production function whereby research labour input is proxied by R&D expenditure. However, when tested empirically, Hingley and Park (2016) find that patent filings are not sensitive to short run movements in R&D , which may reduce performance of the model as a supranational office, the EPO mainly attracts subsequent filings . National offices that receive fewer subsequent filings may find a transfer model unsuitable for forecasting
mainly attracts subsequent filings . National offices that receive fewer subsequent filings may find a transfer model unsuitable for forecasting Hingley and Nicolas (2004) suggest that the EPO could forecast weights for different technology classes of patenting . However, the EPO has previously avoided bottom-up forecasts by technology class on account of finding that these overestimate aggregate filings due to overlap across classes
could forecast weights for different technology classes of patenting . However, the has previously avoided bottom-up forecasts by technology class on account of finding that these overestimate aggregate filings due to overlap across classes low survey response rates, as observed by the EPO in their pre-2021 annual surveys, means findings may be affected by non-response bias. This occurs when respondents to the survey differ systematically to non-respondents. For example, applicants may be more likely to respond to the survey if they have intentions to file patent applications, leading to upward bias in forecasts
in their pre-2021 annual surveys, means findings may be affected by non-response bias. This occurs when respondents to the survey differ systematically to non-respondents. For example, applicants may be more likely to respond to the survey if they have intentions to file patent applications, leading to upward bias in forecasts Dannegger and Hingley (2013) examine the predictive accuracy of patent filing forecasts from annual surveys, using nine years of EPO survey data. They find that accuracy of forecasts based on survey data is highest for the first year after each survey’s base year and declines noticeably for the two following years. This implies that survey methods should not be used for longer term forecasts
European Union Intellectual Property Office (EUIPO)
the EUIPO forecasts registered EU trade mark ( EUTM ) and registered community design ( RCD ) filing volumes, to make decisions on future budget and staff planning. Its forecasting method has changed over the years, from historical trend projection to ARIMA modelling, to Vector Autoregressive ( VAR ) modelling used today. This advancement was largely a response to increased volatility.
) and registered community design ( ) filing volumes, to make decisions on future budget and staff planning. Its forecasting method has changed over the years, from historical trend projection to modelling, to Vector Autoregressive ( ) modelling used today. This advancement was largely a response to increased volatility. between 2011 and 2015, the EUIPO used historical trends to forecast filings, applying an average historical growth rate of ~5% per year. In 2016 and until 2021, the EUIPO used an Autoregressive integrated moving average ( ARIMA ) model to forecast filings . ARIMA models forecast based on moving averages of historic data, to smooth out the influence of outliers and to consider trends, cycles, and seasonality. This method performed well until the shock of COVID-19 in 2020. The office then turned to using a vector autoregressive ( VAR ) model, to incorporate relevant economic variables into forecasting.
) model to forecast filings . models forecast based on moving averages of historic data, to smooth out the influence of outliers and to consider trends, cycles, and seasonality. This method performed well until the shock of COVID-19 in 2020. The office then turned to using a vector autoregressive ( ) model, to incorporate relevant economic variables into forecasting. the EUIPO now uses a VAR model with economic explanatory variables to forecast EUTM and RCD filings on a quarterly basis . This model treats every variable as endogenous, explained by its past values and past values of other variables in the model, allowing dynamic behaviours to be described
model with economic explanatory variables to forecast and filings on a quarterly basis . This model treats every variable as endogenous, explained by its past values and past values of other variables in the model, allowing dynamic behaviours to be described variables were selected for inclusion to the EUIPO’s VAR model by comparing filing trends to National Accounts (NA) indicators for domestic demand, consumption and investment, as well as confidence indicators (based on business and consumer surveys). 10 VAR models were estimated with different combinations of variables to select the one with best fit: a model with 8 endogenous variables and 2 lags. Variables include: EUTM filings, RCD filings, three confidence indicators (for industry, services and consumers), and 3 variables from National Accounts: private FCE (final consumption expenditure), NFCF (net fixed capital formation), and net capital transactions with RoW (rest of world)
model by comparing filing trends to National Accounts (NA) indicators for domestic demand, consumption and investment, as well as confidence indicators (based on business and consumer surveys). 10 models were estimated with different combinations of variables to select the one with best fit: a model with 8 endogenous variables and 2 lags. Variables include: filings, filings, three confidence indicators (for industry, services and consumers), and 3 variables from National Accounts: private FCE (final consumption expenditure), NFCF (net fixed capital formation), and net capital transactions with RoW (rest of world) granger causality tests were performed to better understand the relationship between the variables in the VAR model. EUTM filings were found to be most responsive to confidence indicators in the industry sector and net capital transactions with the rest of the world. Design filings were found most responsive to consumer confidence indicators and private final consumption expenditure
model. filings were found to be most responsive to confidence indicators in the industry sector and net capital transactions with the rest of the world. Design filings were found most responsive to consumer confidence indicators and private final consumption expenditure the EUIPO found their VAR model improved upon ARIMA for forecasting EUTM and RCD filings . The presence of a cointegration relationship among the variables was rejected when tested, suggesting that a VAR model is appropriate, rather than a Vector Error Correction model ( VECM ).
model improved upon for forecasting and filings . The presence of a cointegration relationship among the variables was rejected when tested, suggesting that a model is appropriate, rather than a Vector Error Correction model ( ). to forecast how filings respond and evolve to a shock to one of the economic variables in their VAR model, the EUIPO uses an Impulse Response Function (IPF). A shock to industrial sector confidence is found to have a large effect on EUTM filings in Q1, decreasing in magnitude in Q2, but still holding a cumulative effect after 2 years. A shock to consumer confidence shows greatest impact on RCD filings in Q1 and Q2, and a cumulative impact after 2 years of the same size as the initial effect. This helps forecasters to understand how filings might respond during periods of increased economic volatility
model, the EUIPO uses an Impulse Response Function (IPF). A shock to industrial sector confidence is found to have a large effect on filings in Q1, decreasing in magnitude in Q2, but still holding a cumulative effect after 2 years. A shock to consumer confidence shows greatest impact on filings in Q1 and Q2, and a cumulative impact after 2 years of the same size as the initial effect. This helps forecasters to understand how filings might respond during periods of increased economic volatility typically, three scenarios of forecasts produced by VAR are presented to the EUIPO’s Executive Director, who makes the final decision on which scenario is selected for budget and planning purposes
Gabaly and Hidalgo (2017) evaluate forecasting methods on 1996-2015 EUIPO trade mark and design filing data, finding that VAR models, using explanatory variables such as R&D investment or economic growth, presented an improvement compared to the prediction and modelling power of classic forecasting techniques such as trend projections (using ARIMA ), exponential smoothing or classic econometric methods . Forecasting accuracy was also found to be improved using ITF (Intelligent Transfer Function), an intelligent optimisation algorithm used to select variables, parameters, lags, outliers and optimal transformations to best predict future values. This identified GDP growth as the best predictor of trade mark filing growth, from several economic, IP and business indicators
models, using explanatory variables such as R&D investment or economic growth, presented an improvement compared to the prediction and modelling power of classic forecasting techniques such as trend projections (using ), exponential smoothing or classic econometric methods . Forecasting accuracy was also found to be improved using (Intelligent Transfer Function), an intelligent optimisation algorithm used to select variables, parameters, lags, outliers and optimal transformations to best predict future values. This identified GDP growth as the best predictor of trade mark filing growth, from several economic, IP and business indicators VAR models, as used by the EUIPO, are increasingly being used for forecasting purposes, by the European Central Bank , the International Monetary Fund ( IMF ) , the European Commission’s Directorate General for Economic and Financial Affairs (EC DG-ECFIN) , and the Organisation for Economic Cooperation and Development ( OECD )
models, as used by the EUIPO, are increasingly being used for forecasting purposes, by the European Central Bank , the International Monetary Fund ( ) , the European Commission’s Directorate General for Economic and Financial Affairs (EC DG-ECFIN) , and the Organisation for Economic Cooperation and Development ( ) one difficulty faced when using VAR models is that, as the number of parameters that must be estimated increases, degrees of freedom in the model decrease, which can reduce precision of estimates. It is therefore important to only include variables that hold explanatory power, making tests such as Granger causality, as performed by EUIPO, important. These tests identify a causal relationship between EUTM filings and RCD filings, implying further research is required to understand the interdependence of demand for different IPRs .
models is that, as the number of parameters that must be estimated increases, degrees of freedom in the model decrease, which can reduce precision of estimates. It is therefore important to only include variables that hold explanatory power, making tests such as Granger causality, as performed by EUIPO, important. These tests identify a causal relationship between filings and filings, implying further research is required to understand the interdependence of demand for different . the Impulse Response Function ( IRF ) can be used to model different scenarios in the VAR model, for example identifying the impact on trade mark filings if consumer confidence drops to some level. Other offices may benefit from this to understand the potential impact of shocks.
) can be used to model different scenarios in the model, for example identifying the impact on trade mark filings if consumer confidence drops to some level. Other offices may benefit from this to understand the potential impact of shocks. other IP offices that receive a large proportion of international registrations from WIPO may benefit from using quarterly, as opposed to monthly filing data in their forecasts. EUIPO do so to avoid irregular seasonal and calendar effects that may result from lacking registration date for WIPO filing data
Spanish Patent and Trademark Office ( SPTO )
research has been conducted for the SPTO to develop a methodology which predicts changes in the number of national patent and trade mark applications for a time horizon of three years. Hidalgo and Gabaly (2012) tested various methods on 1979-2009 SPTO data, including an exponential smoothing model (Holt type) , an auto-regressive model of order 1 (AR1), and an ARIMA model. These models were each found to explain at least 50% of the variation in trade mark applications, and at least 80% of variation in patent applications. Taking other metrics for goodness-of-fit into consideration (mean squared errors, Bayesian information criterion), the authors concluded that the ARIMA (1,1,0) model in natural logarithms is the most useful for forecasting, for both patent and trade mark application time series
to develop a methodology which predicts changes in the number of national patent and trade mark applications for a time horizon of three years. Hidalgo and Gabaly (2012) tested various methods on 1979-2009 data, including an exponential smoothing model (Holt type) , an auto-regressive model of order 1 (AR1), and an model. These models were each found to explain at least 50% of the variation in trade mark applications, and at least 80% of variation in patent applications. Taking other metrics for goodness-of-fit into consideration (mean squared errors, Bayesian information criterion), the authors concluded that the (1,1,0) model in natural logarithms is the most useful for forecasting, for both patent and trade mark application time series in a later paper, (Hidalgo & Gabaly, 2013),tested the ARIMA model against a predictive lag variable model, a polynomial distributed lag ( PDL ) model and an intelligent transfer function ( ITF ) model, on 2011-2014 SPTO data. They found all three models improved upon the prediction and modelling power of the ARIMA model, and in particular the ITF model surpasses all other models tested in terms of the degree of fit (fewer errors), improving the selection of predictors and optimizing the predictions which are obtained. Inclusion of GDP and the industrial production index as explanatory variables was also found to improve predictive accuracy
autoregressive models offer predictive accuracy if there is sufficient correlation between movement of current and past filings. However, several studies have found that changes in the series of patents and trade mark applications tend to be largely influenced by milestones and regulatory changes. For example, in the case of Spanish patent applications there was a heavy decrease in the year of 1986 due to the change in Spanish legislation as a result of the enactment in Spain of the Munich European Patents Convention. This causes autoregressive models to suffer from estimation problems, as historical filing trends do not explain future changes caused by idiosyncratic factors
have found that changes in the series of patents and trade mark applications tend to be largely influenced by milestones and regulatory changes. For example, in the case of Spanish patent applications there was a heavy decrease in the year of 1986 due to the change in Spanish legislation as a result of the enactment in Spain of the Munich European Patents Convention. This causes autoregressive models to suffer from estimation problems, as historical filing trends do not explain future changes caused by idiosyncratic factors Hidalgo & Gabaly’s (2013) findings show that predictive accuracy of AR models can be improved by including other explanatory variables (GDP and the industrial production index) and using an ITF model to test the degree of fit for optimisation
Federal Intellectual Property Institute of Switzerland ( IPI )
the IPI recently began using simple trend models and ARIMA models for its IP forecasting, led by a team of 3 economists
recently began using simple trend models and models for its IP forecasting, led by a team of 3 economists other forecasting methods have also been tested on IPI data. Bock et al (2004) applied a structural state-space model to forecast trade mark applications received by IPI between 1992 and 2004. This model uses state, input and output variables to model a time series by a set of first-order differential equations. Explanatory variables, including the Dow Jones index and Swiss consumer confidence index, were also included in the model. It was found to perform better than an ARIMA model with trend, seasonal and random components, based on RMSE (Root Mean Square Error) and relative RMSE indicators. However, it failed to predict “extraordinary situations”, including a steep increase in applications received in the year 2000, associated with a decrease in application fees and the dot-com boom
Hingley and Nicolas (2004) find AR models perform poorly under certain market conditions, such as financial crises or periods of rapid technological change , when IP filings fluctuate greatly from past trends. Similarly, structural state-space models fail to predict changes driven by idiosyncratic factors, though they are found to offer improvement on forecasting relative to ARIMA modelling
models perform poorly under certain market conditions, such as financial crises or periods of rapid technological change , when IP filings fluctuate greatly from past trends. Similarly, structural state-space models fail to predict changes driven by idiosyncratic factors, though they are found to offer improvement on forecasting relative to modelling evidence, based on Swiss data, suggests that inclusion of economic explanatory variables improves IP forecasting accuracy, replicating findings on data at other IP offices ( EPO , EUIPO, SPTO , USPTO )
United States Patent and Trademark Office ( USPTO )
the USPTO uses a simulation tool called the Patent Pendency Model ( PMM ) , implemented in Excel, to predict and simulate patent examination outcomes. The model forecasts using indicators on the supply and demand side. Supply indicators include historical data on manpower at the patent office (patent examiner hires, overtime worked, and examiner attrition rate). Demand indicators include filing forecasts (projections of annual patent application filings based upon forecasts and consensus judgments about the future) and RCE (request for continued examination) Filings. Model outputs are used to predict and simulate future workloads and resources in the examination processes, efficiency and performance
uses a simulation tool called the Patent Pendency Model ( ) , implemented in Excel, to predict and simulate patent examination outcomes. The model forecasts using indicators on the supply and demand side. Supply indicators include historical data on manpower at the patent office (patent examiner hires, overtime worked, and examiner attrition rate). Demand indicators include filing forecasts (projections of annual patent application filings based upon forecasts and consensus judgments about the future) and RCE (request for continued examination) Filings. Model outputs are used to predict and simulate future workloads and resources in the examination processes, efficiency and performance the USPTO forecasts application types separately (original filings, continuing type applications, PCT National Phase, Design filings, and Requests for Continued Examination (RCE), and Provisional filings).
forecasts application types separately (original filings, continuing type applications, National Phase, Design filings, and Requests for Continued Examination (RCE), and Provisional filings). the USPTO considers various econometric models when forecasting original filings. These include economic indicators as covariates, including GDP, unemployment data, and CPI
considers various econometric models when forecasting original filings. These include economic indicators as covariates, including GDP, unemployment data, and CPI predictions of patent renewals (maintenance) utilises linear trend models based on historical patent renewal rates and the number of patents granted in prior years
Trade mark filings are forecast in a similar way to patent filings. Several econometric models are used to predict trademark filings using prior period filings and economic data such as GDP, unemployment, and CPI
future trademark renewals are predicted using a similar methodology as that of patent renewals. A simple linear trend is extrapolated based on historical renewal rates and trademarks registered in prior years
for design patent applications, a simple linear trend model is employed to predict future filing activity
the USPTO considers legislative or legal changes that could affect filings such as fee rate changes for USPTO services and significant changes to patent law/procedures.
considers legislative or legal changes that could affect filings such as fee rate changes for services and significant changes to patent law/procedures. until 2007, the USPTO used the Delphi method to inform forecasts. This is a method of survey data interpretation, whereby a panel of experts adjust their forecasts after each survey round to achieve a consensus forecast for patent filings based on group opinion
the USPTO makes its simulation tool, PPM, available online through an interactive spreadsheet that can be accessed in excel. This may help patent applicants to understand how long their patent pendency period will last, based on supply and demand factors at the IP office.
World Intellectual Property Organisation ( WIPO )
at WIPO , various models are employed to enhance the accuracy of forecasting PCT (Patent Cooperation Treaty) patent filings, Madrid trademark filings, and renewals, as well as Hague design filings and renewals. To forecast filings, a linear trend model and the ARIMA model are used to analyse monthly data. To forecast renewals, a transfer model is employed that leverages yearly registration data and historical renewal percentages across different renewal cycles
, various models are employed to enhance the accuracy of forecasting (Patent Cooperation Treaty) patent filings, Madrid trademark filings, and renewals, as well as Hague design filings and renewals. To forecast filings, a linear trend model and the model are used to analyse monthly data. To forecast renewals, a transfer model is employed that leverages yearly registration data and historical renewal percentages across different renewal cycles in addition to forecasting filing volumes, WIPO also predicts factors that have an impact on revenue, including the number of pages, the number of designs per application, and various fee reductions. By considering these factors, WIPO aims to provide accurate forecasts of revenue generation
also predicts factors that have an impact on revenue, including the number of pages, the number of designs per application, and various fee reductions. By considering these factors, aims to provide accurate forecasts of revenue generation WIPO ’s forecasts are used internally by its finance and operational departments. Typically, financial targets are set below forecasted values, whereas operational targets (such as translation volume) are set slightly higher
WIPO presents its forecasts with a range that is based on an 80% confidence level. This approach allows users (finance and operational departments of WIPO ) to make informed decisions by considering the range of possibilities. Users can then adjust their financial and operational targets, taking into account the forecasting results as well as any associated risk factors.
presents its forecasts with a range that is based on an 80% confidence level. This approach allows users (finance and operational departments of ) to make informed decisions by considering the range of possibilities. Users can then adjust their financial and operational targets, taking into account the forecasting results as well as any associated risk factors. the inclusion of other economic indicators, such as GDP, as predictive variables was initially tested by WIPO , but ultimately discontinued due to the lack of improvement in model performance. WIPO ’s systems provide an alternative filing route, making it challenging to establish a straightforward correlation between changing economic conditions and applicants’ filing strategies.
The European Trade Mark and Design Network ( ETMDN ) / Cooperation Fund project
the European Trade Mark and Design Network ( ETMDN ) connects the EUIPO with national EU IP offices and user associations . In 2013, the network initiated a project to evaluate the best methods for forecasting trade mark and design filings (Havermans, Gabaly and Hidalgo, 2017). The working group consisted of experts from the EUIPO and national IP offices of Denmark, Hungary, Poland, Portugal, Spain, and the UK, with the EPO acting as an observer, and support from the Cooperation Fund .
) connects the EUIPO with national EU IP offices and user associations . In 2013, the network initiated a project to evaluate the best methods for forecasting trade mark and design filings (Havermans, Gabaly and Hidalgo, 2017). The working group consisted of experts from the EUIPO and national IP offices of Denmark, Hungary, Poland, Portugal, Spain, and the UK, with the acting as an observer, and support from the Cooperation Fund . various forecasting methods were tested on EUIPO trade mark and design filings from 1996 to 2015 . Linear regression ( LR ) and AI techniques (support vector machines ( SVM ) and artificial neural networks ( ANN )) were found to outperform traditional forecasting approaches including trend extrapolation, exponential smoothing, and ARIMA techniques. These AI techniques use an algorithm to select the optimal model that best fits the data, employing intelligent optimisation: selecting variables, parameters, lags, outliers and optimal transformations to best predict future values. This was one of the first studies to apply these techniques to the IP-forecasting area.
. Linear regression ( ) and techniques (support vector machines ( ) and artificial neural networks ( )) were found to outperform traditional forecasting approaches including trend extrapolation, exponential smoothing, and techniques. These techniques use an algorithm to select the optimal model that best fits the data, employing intelligent optimisation: selecting variables, parameters, lags, outliers and optimal transformations to best predict future values. This was one of the first studies to apply these techniques to the IP-forecasting area. findings informed the development of a new online forecasting tool, allowing the user to forecast trade mark and design filings by selecting a forecasting model ( SVM , ANN , or LR ), explanatory variables (GDP and/or unemployment growth) and a forecasting scenario (baseline, upside, or downside). The interface can be observed in figure 2
, , or ), explanatory variables (GDP and/or unemployment growth) and a forecasting scenario (baseline, upside, or downside). The interface can be observed in figure 2 the forecasting tool was adopted by 22 EU member state IP offices. However, according to an EUIPO source, no office has yet used the model to forecast for budgeting purposes
AI pattern recognition and self-improving capabilities can improve the accuracy of forecasts compared to traditional methods. However, AI techniques may be regarded by some as a ‘black box’ approach. Algorithms can reach a high internal complexity in computational statistics terms, which can make understanding and interpretation by analysts difficult
pattern recognition and self-improving capabilities can improve the accuracy of forecasts compared to traditional methods. However, techniques may be regarded by some as a ‘black box’ approach. Algorithms can reach a high internal complexity in computational statistics terms, which can make understanding and interpretation by analysts difficult AI forecasting does not require expert intervention for parameterisation and selection, as other modelling techniques do. This is because they are algorithms that are mainly based on automatable machine learning techniques, which automatically take the relevant decisions to improve the forecasting process. However, lack of human analytical understanding can reduce interpretability of outputs
forecasting does not require expert intervention for parameterisation and selection, as other modelling techniques do. This is because they are algorithms that are mainly based on automatable machine learning techniques, which automatically take the relevant decisions to improve the forecasting process. However, lack of human analytical understanding can reduce interpretability of outputs for 2015, the SVM model predicted 108,028 EUIPO trade mark filings, compared to an actual of 108,028, giving an error of 0.4%. The average error across the three models ( SVM , ANN and LR ) in forecasting 2013-2015 trade mark and design filings was 1.6% and 2.6% respectively. In no case did one of the models consistently outperform the other three across all forecasting years (see Figures 3 and 4)
4. Approaches to forecasting: Other selected organisations
Organisation Forecasting method(s) used Thomson Reuters * Trend extrapolation Patent Forecast * AI -based interactive visualisation
* Analyst intelligence Department for Transport * Error correction model ( ECM ) National Grid / Ofgem and energy sector * Scenario-based forecasting
* Support Vector Regression (SVR) Eurostat * Time-series factor model OECD * Indicator models
* Structural change analysis Euromonitor * Scenario-based forecasting
* Industry forecast model based on elasticities
Thomson Reuters is one of the few private organisations found to forecast patent filings. It uses extrapolation techniques based on average growth rates over the last five years to forecast future patent filing volumes for economic areas (including China, the EU, Japan, South Korea and the US) and uses this information to calculate total and domestic filing growth rates.
Thomson Reuters last updated their patent filing forecasts in 2010, according to publicly available information. Forecasting methods have since advanced beyond trend extrapolation, to forecast fluctuations in filings from their historical trend
Patent Forecast is an independent company that offers interactive visualisation software of AI -sourced patent and market data, updated every week across 56 sectors. It uses a combination of AI and employee analysts to spot market shifts and industry trends in patent filings using its proprietary software. This evidence informs sector-specific forecasts published in articles on its website.
using AI to identify market shifts and industry trends may provide valuable anecdotal evidence to accompany forecasting models, in addition to evidence gathered from expert panels
to identify market shifts and industry trends may provide valuable anecdotal evidence to accompany forecasting models, in addition to evidence gathered from expert panels speculation and use of anecdotal evidence reduces ease of forecasting relative to simplified models that can be clearly defined by their assumptions.
The Department for Transport ( DfT ) runs the “National Air Passenger Demand Model (NAPDM)” , which forecasts passenger demand for air travel out to 2050 . An Error Correction Model ( ECM ) is used to estimate elasticity of demand for air travel with respect to income, price (air fare), and measures of economic activity (such as GDP). This tells us how responsive demand is to a change in one of these variables
) runs the “National Air Passenger Demand Model (NAPDM)” , which forecasts passenger demand for air travel out to 2050 . An Error Correction Model ( ) is used to estimate elasticity of demand for air travel with respect to income, price (air fare), and measures of economic activity (such as GDP). This tells us how responsive demand is to a change in one of these variables an ECM is used when dealing with cointegrated data, meaning that two time series (for example, air travel demand and consumer income) have a long-run relationship. Other examples of cointegrated relationships include stock prices and dividends, and consumption and income. The cointegration coefficient estimated within the ECM is used to estimate how long it might take air travel demand to return to its long-term trend following a short-term shock, for example to consumer income. Different lag structures can be tested to remove potential autocorrelation, which occurs when the error terms in the regression are not independent of one another
for an ECM to be used to forecast elasticities with respect to IP filings, a long-run relationship (cointegration) must be found between IP filings and some other time series variable. Josheski and Koteski (2011) find a positive relationship in the long run between quarterly growth of patents and quarterly GDP growth, using the ARDL bounds test . This suggests the suitability of ECM for estimating elasticity of patent filings growth to a change in GDP
to be used to forecast elasticities with respect to IP filings, a long-run relationship (cointegration) must be found between IP filings and some other time series variable. Josheski and Koteski (2011) find a positive relationship in the long run between quarterly growth of patents and quarterly GDP growth, using the ARDL bounds test . This suggests the suitability of for estimating elasticity of patent filings growth to a change in GDP ECM estimates how long it will take a variable (e.g. patent filings) to return to its long-term relationship with another variable (e.g. GDP) from a disequilibrium position caused by some shock. Unlike most other models discussed in this paper, this model addresses forecasting accuracy in the face of idiosyncratic short term shocks
estimates how long it will take a variable (e.g. patent filings) to return to its long-term relationship with another variable (e.g. GDP) from a disequilibrium position caused by some shock. Unlike most other models discussed in this paper, this model addresses forecasting accuracy in the face of idiosyncratic short term shocks The ECM model can be adapted to account for structural breaks in the long-run relationship between the cointegrating variables. A dummy variable is included in the regression model that takes the value of one after the year of the structural break, and zero otherwise.
National Grid / Ofgem and energy sector
the National Grid is responsible for publishing energy demand forecasts under Ofgem’s rules. Accurate forecasts are critical for efficiency; underestimation of energy consumption can lead to power outage, and overestimation can lead to unused capacity and waste of capacity
previously, National Grid produced a single forecast of annual gas demand based on analysis of history and views of the future incorporating forecasts of economic growth, industry intelligence about new developments, new technologies and new connections to the gas network. More recently, National Grid adopted a scenario-based approach to forecasting. Scenario 1 (“Gone Green”) uses a bottom-up approach to calculate energy consumption consistent with renewable energy targets and CO2 reduction targets being met. Scenario 2 (“Slow Progression”) uses econometric modelling, forecasting demand based on a range of factors, including fuel prices, economic growth and number of households
Abdelkader et al (2015) investigate the use of support vector regression (SVR) model to forecast energy demand. This is a machine learning model that solves non-linear optimisation problems to identify the line of best fit. It differs from other machine learning models by looking at the extremes of datasets to draw a decision boundary. The decision model of SVR can be easily updated for new information, and prediction accuracy is found to be higher than other machine learning models (Zhang et al 2018).
the National Grid’s forecasting methods show how scenario-based forecasting can be advanced through use of econometric techniques within scenarios
use of machine learning models for forecasting (such as SVR) have proven successful in the energy sector, and so could be considered for forecasting IP filings
Eurostat, the statistical office of the European Union, forecasts the growth rate of unemployment, GDP and inflation using a time-series factor model, where the factors include macroeconomic variables of monthly and quarterly frequency. An EM (Expectation Maximization) algorithm is used to estimate the common factors and the factor loadings (the weights and correlations between each variable and the factor)
Factor model techniques could be considered for forecasting IP filings, where relevant factors could be identified using machine learning practice, such as expectation maximization algorithm, as used by Eurostat.
OECD , an intergovernmental economic organisation with 38 member countries, forecasts quarterly GDP growth using indictor models . These combine information from both “soft” indicators, such as business sentiment and consumer surveys, and “hard” indicators, such as industrial production, retail sales, house prices etc. and use is made of different frequencies of data and a variety of estimation techniques
, an intergovernmental economic organisation with 38 member countries, forecasts quarterly GDP growth using indictor models . These combine information from both “soft” indicators, such as business sentiment and consumer surveys, and “hard” indicators, such as industrial production, retail sales, house prices etc. and use is made of different frequencies of data and a variety of estimation techniques OECD has also investigated use of machine learning to identify non-linearities and structural change in macroeconomic data, to make forecasts. OECD find that a model with a functional form incorporating non-linearities (multiple interactions, discontinuities and structural breaks) performs better than a simple AR (1) benchmark model ( OECD , 2020). Non-linearities are common in macroeconomic data. For example, growing house prices may signal strong GDP growth up until a given threshold, beyond which the bubble bursts and the economy may decelerate. The method of structural change analysis aims to address these challenges
combining quantitative and qualitative forecasting methods reduces reliance on survey data, which has been shown to have a short horizon of predictive accuracy.
Euromonitor, a UK market research company, forecasts GDP and inflation using a macroeconomic model covering a number of scenarios. Category market sizes are projected by forecasting upside or downside pressures using income and price elasticities. Final forecasts are further refined by industry experts to account for hard-to-predict events, such as legislation changes or key company marketing campaigns
Euromonitor’s industry forecast model identifies “driver effects” by industry. The model calculates elasticity of retail volume sales in specific industries to a change in GDP per capita, product price, habit persistence and population growth, based on historical time series data. These drivers are then used for forecasting (in a similar way to use of factors in a factor model).
identifies “driver effects” by industry. The model calculates elasticity of retail volume sales in specific industries to a change in GDP per capita, product price, habit persistence and population growth, based on historical time series data. These drivers are then used for forecasting (in a similar way to use of factors in a factor model). figure 5 outlines Euromonitor’s approach to forecasting
use of income and price elasticities to forecast upside and downside pressures could be replicated for forecasting IP filings, if these are found to be sensitive to customer changes in income and fees
there is evidence of private companies using advanced forecasting methods to forecast product sales. Researchers at Sun Microsystems, a US computer manufacturer now owned by Oracle Corporation, use a Bayesian dynamic linear mixture model to forecast their product sales. This is a machine learning model that uses unknown quantities estimated from noisy measurements. The state of the system evolves from one state (time t) to another (time t+1) according to a known linear transition equation, which can include random disturbances and intervention effects. State values are successively predicted given the knowledge of the past observations, and then updated upon the reception of the next observation.
Yelland and Lee (2003) find higher forecasting accuracy when using dynamic linear models compared to exponential smoothing. However, processing speed of machines running DLMs can be very long, as exhaustive examination of all possible model sequences for all periods is required. Therefore, use of DLMs may prove difficult for large datasets, such as historical IP data
Figure 1: European Patent Office transfer model by bloc
Figure 2: TMDN trade mark and forecasting tool user interface
Figure 3: Results of use of the TMDN trade mark and forecasting tool to forecast EUIPO trade mark filings
Figure 4: Results of use of the TMDN trade mark and forecasting tool to forecast EUIPO design filings
Figure 5: Illustration from Euromonitor website showing its process for updating macroeconomic forecasts for external events.
Disclaimer
This article is intended for general information purposes only and does not constitute legal advice. For advice specific to your situation, please contact our team at T & M Legis for a consultation with our Legal Experts.
