Trend analysis as a starting point for making long-term projections
The choice for an appropriate projection method depends on the level of uncertainty (see strategy page for discussion on four levels of uncertainty defined by prof. Walker). Trends & Transport offers strategic advice in developing (long-term) projections at each level of uncertainty. Analysis of trends and key uncertainties acts as a starting point for assessing the level of uncertainty, the appropriate projection methods, and the suitable policy methods. Future developments that can be adequately described by means of trends and plausible uncertainties can still be handled with policy methods for dealing with Level 3 uncertainty. This implies that one does not have to take a broad range of possible scenarios into account and can base the policy on a more limited set of plausible scenarios. Trend analysis also provides valuable input when developing long-term forecasts, because forecasts will only remain valid as long as the underlaying system of trends remains stable. Trend analysis is thus a starting point for assessing the level of uncertainty and making long-term projections. Below follows a discussion of recommended projection methods at each level of uncertainty. The section on Level 3 uncertainty presents a new integrated framework for assessing trends and key uncertainties that I developed over the past few years.
Forecasting techniques
Ex-post forecast with an ARIMA model
Source: own calculations with Python for ARIMA (1,1,0) model.
Ex-post forecast with a causal model
Deterministic forecasting (Level 1)
When forecasting just a few periods ahead, use can be made of auto regressive integrated moving average models (ARIMA) that are based on historical data of the trend itself and do not require any input from exogenous explanatory variables. Auto regressive models are easy to implement and even provide Level 2 projection intervals if desired. However, they are less suitable for making long-term projections, because the model variables are all linked to lagged values of the system itself, for which the forecast is still based on mere trend extrapolation. This implies that the forecast only remains valid as long as the underlaying trend and variance remain stable, which is questionable on the long run. When the underlaying trend changes the lack of causal relations to exogenous variables makes the model unable to adjust. For long-term projections it is recommended to use causal relations to exogenous variables.
The upper figure shows an ex-post ARIMA model based on historical data from 1948 to 1970, for which the actual trend deviates from the projection, but more or less remains within the 95% confidence interval. The lower figure shows a similar throughput projection based on the historical relation between GDP and port throughput in the period from 1948 to 1970, and perfect foresight into the GDP from 1970 to 2010. Different causal relations were tested. The trend almost perfectly followed the forecasted values for the differences approach. For the other approaches the statistical properties indicated a problem with an autocorrelated error-term, which in principle disqualifies the use of these models for forecasting purposes.
In practice long-term transport projections are based on a top-down, a bottom-up, or a logistical modelling approach. The top-down approach start with a macroeconomic assessment of the hinterland and estimate the transport volumes on the basis of causal relations to exogenous variables such as population, GDP, and trade volumes. The bottom-up approach analyses planned micro or meso-economic developments for individual companies or industrial sectors, which are aggregated into a total estimate for the region. The logistical modelling approach is applied when transport flows are footloose and not related to a specific hinterland area, such as for transfer passengers on an airport or for transhipment containers in a seaport. Logistical models are used to assign transport volumes over a larger network.
Probabilistic forecast
Different views on economic growth
Probabilistic projection with realisation
Source: subset of projection from van Dorsser, C., Wolters, M. A., & van Wee, B. (2012). A very long term forecast of the port throughput in the le-Havre – Hamburg range up to 2100. European Journal of Transport and Infrastructure Research, 12(1), 88–110 , data extended with data on realised throughput volumes from published port statistics.
Probabilistic forecasting (Level 2)
Probabilistic forecasts can be based on similar causal forecast models as for Level 1 uncertainty, but require a priori statements on the distribution, mean and variance of the parameters applied in the model. The notion of these uncertainties can either be frequentist or Bayesian. In the frequentist notion the uncertainties refer to the frequency of occurrence whilst in the Bayesian notion probability is also regarded as a degree of personal belief and therefore subjective, unique and subject to change. The latter can for instance be based on expert-judgement. Probabilistic forecasts require simulation and can be combined with financial models to define not only the volumes, but also the profitability in terms of probabilities. A prerequisite for the use of probabilistic forecasts is that probabilities can be assigned to all uncertain parameters, if this is not the case for some of the key uncertainties, one has to shift towards the use of Level 3 or Level 4 approaches and develop scenarios (or probabilistic scenarios). In case of Level 4 uncertainty, forecasting models can still be useful to quantify scenarios.
Simple probabilistic forecasts can be obtained by applying ARIMA models (or SARIMA when including seasonal effects). However, such autoregressive models that are solely based on the own historical movement of the trend and that do not include any exogenous explanatory variables are considered less suitable to forecast a long period ahead. Long-term forecasts are preferably based on causal relation to exogenous variables. These variables need to be better forecastable than the concerned variable itself, such as for instance for population and GDP.
To illustrate the use of probabilistic models. In 2009 I prepared a probabilistic long-term forecast for the total port throughput in the Le Havre–Hamburg range up to 2100, that got published in 2012. The forecast uses a combination of System Dynamic Modelling, Expert Judgement, and Causal Relations. On the basis of the forecast it can be expected that the port throughput in the Le Havre–Hamburg region will remain growing throughout the first half of the century, but at a reduced pace.
An essential aspect of the forecast is the probabilistic GDP projection, which I developed myself based on trends in population, labour population, hours worked, and labour productivity per hour. At this point there is major issue with the mainstream neoclassical assumption on labour productivity, for which I proposed to apply and alternative paradigm. Until now, the realisation of the forecast is still in line with the median estimate. It would however been far off if I would have applied the mainstream projections. I therefore recommend developing your own GDP projections, or at least take a critical attitude towards mainstream economic growth projections.
Framework of trends
Port METATRENDS report (442 pages)
Three-layered meta-framework
Source: based on
van Dorsser, C., Taneja, P., & Vellinga, T. (2018) PORT METATRENDS, Impact of long term trends on business activities, spatial use and maritime infrastructure requirements in the Port of Rotterdam, TU Delft and Smart-Port
and
van Dorsser, C. and Taneja, P. (2020) An integrated three-layered foresight framework.
.
Trends and key uncertainties (Level 3)
Since 2017, I have, together with my colleagues at the TU Delft, developed an integrated foresight framework and method to support decision makers who are confronted with today’s complex and rapidly changing world. The method aims at reducing the degree of uncertainty by addressing the inertia or duration of unfolding trends and by placing individual trends in a broader context of a framework of trends. It analyses three layers of trends with a different inertia and starts with a broad systematic identification of relevant megatrends and structural uncertainties. The identified trends are filtered and distributed into three distinct levels, each representing a different level of inertia. Trends that have already existed for more than a century are the most certain. For these trends the direction is relatively clear, and a reasonable estimate for the next 20 years can be made. The following centuries-long trends were identified:
-
Secularisation and individualisation;
-
Nature of activities and social power;
-
Population growth;
-
Urbanization;
-
Energy and raw material use;
-
Technological progress and economic output;
-
Connectivity and information exchange;
-
Climate change and environmental degradation;
-
Transport costs and globalization;
-
Shifts in geopolitical world order.
Relatively certain are also the pervasive drivers of the so-called Kondratieff waves. These waves reveal an about 50 years lasting cyclical movement in the world economy. The world is now at the end of the 5th Kondratieff wave that is driven by globalization and ICT and gradually moving towards the 6th Kondratieff wave that is developing around sustainability and the internet of things (IoT). It is expected that the next 20 years will be dominated by innovation and the development of new sustainable- and data driven technologies and business models. Around the year 2040, these new activities will presumably take over the dominant position in the business models and provide a solid basis for a new sustainable economy.
Storyline scenarios
Source: own representation based on Davis (2002) Scenarios as a Tool for the 21st Century, Shell International Ltd.
S toryline scenarios (Level 4)
In case of multiple key uncertainties or no clue at all, it is logical to shift towards the use of a multiplicity of possible scenarios, which may include a number of wildcard scenarios to deal with inevitable surprises. In his book "The art of the long view, panning for the future in an uncertain world" , Schwartz (1991, p.4), explains that “Scenarios are a tool for helping us to take a long view in a world of great uncertainty. The name comes from the theatrical term ‘scenario’ – the script for a film or play. Scenarios are stories about the way the world might turn out tomorrow, stories that can help us recognize and adapt to changing aspects of our present environment. […] Scenarios are not predictions. It is simply not possible to predict the future with certainty" . This description of scenarios relates to the pure use of inductive storyline scenarios to prepare for changes ahead, such as the ones that helped Royal Dutch Shell to take a leading position after anticipating the possibility of an oil crisis.
In practice scenarios are often also used to assess a bandwidth of possible developments that are considered to lay within the realm of normal expectations (i.e. excluding wildcard scenarios). When scenarios are used to illustrate the bandwidth of possible developments, they are often placed in a two dimensional framework in which they are developed around four quadrants that are created by the intersection of the axes of two uncertain key drivers. This kind of scenarios is referred to as deductive scenarios.
Though frequently applied, there are some drawbacks with the use of deductive scenarios, especially when more than two key drivers should have been taken into account. This was for instance the case in 2012, when I was responsible for developing the Shipping Scenarios for the Dutch Delta Programme. The Delta scenarios were developed around the key drivers of economic growth and climate change, whereas I think a third driver reflecting the speed of the transition towards a sustainable society (i.e. the key driver of the 6th K-wave) should have also been included. To resolve this issue, I placed the scenarios in a 3D frame and added two additional scenarios as a sensitivity analysis.
Another drawback of the use of deductive scenarios is that it makes storylines weaker and less thought out than inductive scenarios. They essentially become a fill-inn exercise for the predefined states of the key variables on the axes. This makes them less intuitive and less valuable to prepare an organisation for possible changes ahead. Moreover, they also become boring to read. An interesting intermediate solution is to develop inductive scenarios that are intended to cover the bandwidth of what may be expected on the basis of a well-defined set of deductive scenarios. In this case the storyline remains stronger, while keeping the desired property of illustrating the bandwidth of normal expectations. An example of this approach are the scenarios of the Port of Rotterdam .