1. Introduction
China suffered direct economic losses of more than USD 42 billion during the first nine months of 2023 from natural disasters, including torrential rains, landslides, hailstorms, and typhoons according to government data. This figure was presented at the last COP28 summit during last November–December 2023 jointly with many others coming from different parts of the planet, as reported in recent posts from the World Economic Forum [
1]. In Africa, disasters from 1970 to 2021 caused USD 43 billion in economic losses, according to the World Meteorological Organisation. Europe also reported a cost of around USD 562 billion in losses. For South America, the losses amounted to USD 115.2 billion and for North America, Central America, and the Caribbean it was USD 2 trillion. Meanwhile, the latest instalment of the US National Climate Assessment has concluded that extreme weather events currently cost the country USD 1 billion every three weeks and averaged USD 150 billion in damages each year between 2018 and 2022. All in all, extreme weather, climate- and water-related events caused almost USD 1.5 trillion of economic losses in the last decade according to the World Meteorological Organization. The figure is much higher compared to the USD 184 billion in losses reported during the 1970s.
These costs are expected to become even higher in the future, deeply influencing countries’ GDP [
2]. This might be a reason for each country to take immediate action, mitigate climate change, and avoid future larger damages. However, a careful evaluation of the future economic impact of global warming is required and future costs against present costs must be balanced under some economic assumptions and climate scenarios [
2,
3,
4]. The most basic assumption considers an exponential with a constant discount rate. The choice of discount rate is part of current economic debates on the urgency of the response to global warming (see [
2,
3,
4,
5,
6,
7,
8,
9,
10]). Most normative approaches attempt to derive the discount from axiomatic principles of justice, or from utility theory and assumptions about growth [
11,
12,
13]. Discounting includes impatience, economic growth, and declining marginal utility and it can be embedded in the Ramsey formula [
14,
15,
16]. This approach, however, makes it difficult to make quantitative estimates because the factors involved are difficult to measure empirically.
Here, we take a positive approach, assuming that discounting is equivalent to bond pricing. However, the time to maturity of bonds is not generally higher than 30 years, but climate change time horizons need the discount 100 years or even more into the future. Thus, inferring long maturity bond prices from empirical data on short maturity bonds is fundamental and long historical time series are needed for reliable statistical inference. These data are most often not available and it will be required to consider a reasonable model for real interest rates at different maturities.
There are two effects influencing long-term rates that must be considered. The first of these is risk aversion. Longer-term bonds bear greater risk than shorter ones, and higher risk should imply higher interest rates. The second effect is more subtle, and is due to the fact that interest rates are uncertain and highly persistent. In the environmental context, this effect was originally pointed out by Weitzman [
17] and Gollier et al. [
18] (see also [
19]). However, the effect was also pointed out in a general context in [
20], and it has been known much longer in the context of bond pricing [
21]. The long-run rate is, thus, dominated by the lowest value, since asymptotically all the other discount factors will be negligible in comparison. To see this concretely, consider two possible rates,
and
, with
, and assume they have equal probability. Then, the average discount factor at time
t in the future is
. Note that, since the sum of two exponentials is not an exponential, the discounting function is no longer an exponential. But for
t which is sufficiently large,
, i.e., the discount function becomes approximately exponential with the lower interest rate. This illustrates that when interest rates are uncertain and persistent, lower interest rates tend to dominate.
We build here on earlier empirical work. Stochastic interest rates provided by Litterman et al. [
22], Newell and Pizer [
23], and Groom et al. [
24] are more realistic processes than those provided by Weitzman and Gollier. Groom et al. [
24] also noted that the drop in rates depends on uncertainty and persistence. All these authors calibrated the models suggested against long historical time series of 10-year bonds. We here add to these previous works in several ways. We study century-long records of the prices of 3-month bonds, 10-year bonds, and inflation, for both the US and the UK. Real interest rates are found to become negative more than one third of the time, often by substantial amounts. To our knowledge only a few authors have addressed this issue (e.g., Freeman et al. [
25]), while the usual approach has been to ignore this fact, and instead has forced historical interest rates to be positive in order to be consistent with the proposed models.
We instead take the view that negative real interest rates are a fact that cannot be ignored. This leads us to choose the Ornstein–Uhlenbeck (OU) model, which is compatible with negative real interest rates. The model provides real interest rates that can be negative but they are constantly pulled back to a certain rate. The formula for the yield curve for the OU model has been derived in the finance literature using several approaches (see, for instance, [
26,
27,
28], and the review in [
29]). We derive this in a somewhat simpler way using Fourier transform methods. The previous work in environmental discounting cited above [
24] was based on numerical simulations. In contrast, we take advantage of the existence of closed-form solutions, which allows us to better estimate the long-term discount rate. When the OU model is used in the finance literature, the mean reversion parameter is taken to be so strong that the nominal interest rate goes negative with a probability close to zero. In a different approach, Davidson et al. [
30] examined a square root Ornstein–Uhlenbeck model which can never go negative and which is pulled down toward 0 as it is an absorbing state. Davidson et al. [
30] solved the asymptotic (long) rate by using the Feynman–Kac functional, which is quite different from our approach.
The prevalence of negative real rates makes many of the standard nominal interest rate models inappropriate vehicles for studying real rates, and the OU becomes an obvious choice while preserving the highest possible simplicity. Other models might also be reasonable choices but they would, in any case, need to contain key features, such as mean reversion, random fluctuations, and a normal level of the process [
27]. The paucity of data records available (hundreds of points) makes parameter estimation challenging. Among other possible candidates with the fewest number of parameters and which are analytically tractable with the methodology herein presented, it is possible to consider the shifted Feller model and the log-normal model as we have done elsewhere [
31]. The shift avoids negative rates, but it becomes somewhat artificial and context dependent as it totally depends on the period analyzed [
31]. Another alternative is to add more timescales into the time series and to consider a fast and a slow mean reversion with two different normal levels. This approach can be modeled with a two-dimensional OU process and enables moving forward the stationary limit beyond the data range available [
32]. This sophistication already involves the addition of two more parameters, which are difficult to estimate due to the scarce data available [
31,
32].
Therefore, our main original contribution here regards properly taking risk aversion into account with the OU model. The works described above [
23,
24] were calibrated against a single maturity bond (10 years). Risk aversion is a well-accepted notion in bond pricing, and one would expect models that neglect it to underestimate long-term interest rates. We fix this by taking risk aversion into account and fit the resulting model to both 3-month and 10-year bond yields, which provides us with two points on the yield curve (one for very short maturities, thus capturing instantaneous interest rates, and another bringing out the long-term effects of the market price of risk). This approach can become more precise by taking additional bond maturities or even additional data sources to better account for risk aversion but, to avoid further sophistication in the whole methodology, we have discarded this option.
We here make the hypothesis that the long-term, risk-adjusted, bond discount rate is a proxy for measuring the urgency of climate action, the latter becoming higher as the former becomes lower, and vice versa. For this reason, we herein present a complete empirical study on real interest rates for the US and UK for which we have century-long records, and that are considered amongst the most economically stable countries as far as inflation and economic growth are concerned. Our focus is based on the interdisciplinary approach of complex systems developments. We have also presented elsewhere a similar study for other less stable countries, but with lower statistical reliability due to the scarcity of data and without considering risk aversion (i.e., the market price of risk) [
31]. Several different theoretical aspects of discounting have recently been discussed by some of us, mainly regarding the effects on the discounting of extreme market situations [
33] or the introduction of resettings in real interest rates [
34] (see also the recent review [
35]). These are, however, sophisticated models with a larger number of parameters, thus making model calibration more difficult. This is the reason why we here restrict the analysis to the OU process.
After
Section 1,
Section 2 presents the formal mathematical framework. It introduces the Ornstein–Uhlenbeck model and the addition of risk aversion in its simplest form for the discount function and the long-term interest rate.
Section 3 focuses on the empirical estimates of the real interest rates by taking US and UK empirical data and provides an empirical analysis. It also includes the parameter estimation procedure for the OU model with the inclusion of the market price of risk. The main goal of the section is to estimate the discount function and the long-term interest rates. Their confidence intervals are also obtained.
Section 4 offers a broader interpretation of the results obtained, focuses on some relevant aspects, such as the prevalence of negative rates, and points out some limitations of the current research, alongside consideration of future research directions.
Section 5 summarizes the main results and highlights the practical relevance of the results obtained. Several technical details and descriptions are provided in three different appendices.
4. Discussion
Climate change and climate action are widely studied from a variety of perspectives [
41]. However, researchers in financial economics have only recently turned their attention to climate change, and climate finance is currently a quickly growing research field [
42]. Finance academics participating in a recent survey have identified discount as a key topic to reduce climate risks [
43]. With this paper, we have taken a multidisciplinary perspective from complex systems science and its related methods to study historical bond prices [
44], with the aim to contribute to the need of new approaches to evaluate climate action urgency [
45,
46].
More specifically, we have wanted to highlight that long-term discount rates are not just a trivial matter of extrapolating mean interest rates, but rather that one must take several non-trivial factors into account. To begin with, because real interest rates are so often substantially negative, one should use a model that permits negative rates. This leads us to the Ornstein–Uhlenbeck model. While the presence of negative rates in this model may be viewed as a liability for describing nominal rates, for real rates, this becomes a virtue. Another factor that should be considered is the market price of risk, which tends to raise longer-term rates. Finally, one should properly take into account the uncertainty and persistence of interest rates, which tends to lower the long-term discount rate. The use of the OU model accommodates all of these factors. When we estimate the OU model and compare it with real data, we see good agreement on several essential properties, such as the frequencies of negative rates and yield curve inversions.
Our results indicate that the long-term interest rate used by Stern [
2] is supported by historical data. His value of
% is less than a standard deviation below the estimated long-term rate for the UK of
%, and just under two standard deviations of the US long-term rate of
%. More recent estimates by Stern indeed suggest a scenario with a value below 1% [
45]. In contrast, higher long-run rates—see, for instance [
6,
7]—seem not to be supported, as they are well above the 95% confidence intervals of
for the UK and
for the US. Our estimates of
% (UK) and
% (US) are compatible with the rates recently estimated by Giglio et al. ([
47,
48,
49]), which use data from UK housing markets during 2004–2013 and Singapore during 1995–2013 to estimate an annual discount rate of 2.6% for payments more than 100 years in the future.
The prevalence of negative rates and their role in the whole analysis merit an even deeper analysis than the one provided here so far. We have already explored the role of negative rates in a previous publication, which, however, did not incorporate the market price of the risk component [
31]. By still considering the OU model as a valid model, one can then see that larger fluctuations (and, therefore, a higher probability to have negative rates) directly implies even lower long-term rates. This effect can indeed also be observed in Equation (
22), where the term related to the noise amplitude,
, has a negative sign. This effect is magnified in less stable countries compared to the UK and US [
31]. The results from countries with a higher prevalence of negative rates and their even lower (and eventually negative) long-term interest rates further increase the urgency for taking action today rather than tomorrow. A more careful analysis would, however, need to be performed to enable a more quantitative discussion.
Our results could potentially also be improved in other ways, which, in turn, can lead to relevant future research avenues. One new research avenue would be to acquire more data. A broader exploration would enhance the applicability and robustness of the methods and the results being presented here. This could include data from more countries—as we undertook recently, but only for 10-year bonds and, therefore, unable to include risk aversion [
31]. The 3-month data from other countries would allow us to incorporate the market price of risk into the analysis and we would be able to elaborate a more complete study, which indeed could include a rigorous discussion on the sensitivity of the different components that influence the discount function and the long-term interest rate
.
One should, however, be aware that the methodology depends on long time series within which the risk premium evolves through time and within which many structural changes may occur. This certainly may affect the whole analysis, and in practical terms can deeply influence the results based on today’s environment. These aspects should be included if we seek concrete applicability of the results obtained here. Along the same lines, the approach might not be valid for emerging markets and for some countries (climate change has different impacts on different countries), either because they may not have available the data being analysed here or because some countries/markets are far from stable.
In relation to the possibility of adding new data sources, it could also be of interest to include longer term bonds, which would certainly allow the inclusion of a more sophisticated market price of risk function. We have here limited the analysis to a constant
q market price of risk independent of
r and this is another limitation of the current work. Credit spreads data or supply/demand effects in bond markets, inflation expectations, and inflation-indexed bonds would be other possible datasets that could be used to provide a better estimation of the long-term interest ratio. The analysis would eventually need a more sophisticated methodology than the one presented in
Section 3.
Another possible improvement would be to extend the model to better capture the non-stationarity and/or heavy-tailed behavior observed in the data [
36,
50]. In this regard, we suspect that had we been able to take the observed non-stationarity [
45] and/or heavy tails into account, the mean values would have decreased because of the boosting of the uncertainty/persistence effect. This possibility is under investigation presently but we may anticipate that the result could suggest even lower discount rates [
32].
At a broader level, both Stern and Stiglitz have recently provided new methods and models which could be further studied from a complex systems science perspective [
46]. Stern [
45] has indeed put the emphasis on incorporating a stronger multidisciplinary perspective to consider the social and behavioral dimensions [
51,
52,
53]. The same author [
2] also mentions a definition of social discounts that might include inputs from social dilemmas and predefined behavioral experiments [
54].