Data-based Price Forecasts and Scenarios for Recyclates

Imagine you are at the gas station or the stock exchange: If you knew that the price of gasoline or a targeted stock would decrease in the near future, the best strategy would be to delay a purchase lone as possible. If price forecasts pointed upwards, an immediate purchase would be worth it. The same applies to the recyclate sector: The profitability of producing, for example, resource-saving packaging film depends not only on the price of the selected or prescribed material type but also on how well you can estimate the price development of the recycled plastic in advance and purchase accordingly. But how do you get the assurance that you are making the right purchasing decision? Can sound, data-based forecasts of the future price of regranulates shed some light into the darkness? Yes, they can! We explain how this works using the example of forecasting recyclate prices in May 2024 and using the forecasting software future.

The basis for price forecasts: the time series

Our goal is to generate forecasts for recyclate prices that are included monthly in the Recycario report by Peter Jetzer on the Kreislaufgold blog . For this, we first look at the data basis available to us. Since we are creating the price forecast for the coming months and want to include influencing factors that are also available on a monthly basis, we make sure that we can work with historical recyclate prices on a monthly basis. In our case, aggregating the average price for recyclates per month ensures a gapless, well-maintained monthly time series. Since we are dealing with a time series starting in 2013, another basic prerequisite – that the time series is sufficiently long – is satisfied. The most recent value of the average recyclate price available to us belongs to May 2024. Accordingly, we can create forecasts for June 2024, July 2024, etc.

Price time series often tend to have a “smooth” pattern

Initially, we create forecasts based solely on historical recyclate prices to obtain a benchmark for an attainable forecast accuracy. We can include calendrical factors such as seasonality or the trend of the time series. The recyclate prices do not show any visible or verifiable seasonality, as is often the case with prices in general. The course of recyclate prices is essentially driven by autoregressive behavior, i.e., the prices of previous months have a strong influence on the prices of the current month, which is reflected in a certain trend behavior of the time series. Other, potentially less easily visible patterns are not verifiable in our case. We simulate the historical forecast accuracy and find that we achieve a mean absolute forecast error of about 1.9% for the 1-step forecast – that is, for the price of the next month.

Figure 1: Benchmark forecast of recyclate prices based solely on historical recyclate prices.

Appropriate additional information should improve the forecast model

However, the forecast accuracy that we can achieve by using only historical recyclate prices is not sufficient for us. Not least because there is a strong suspicion that among the multitude of external influencing factors, there may be driving factors for the price developments lurking. In discussion are, among other things, prices of primary plastics, prices of their waste, as well as general economic indicators. It is also easy to imagine that political measures influence the price developments of regranulates. The following figure provides an overview of possible drivers from the economic/cyclical, market-specific/technical, ecological, and regulatory perspectives:

Figure 2: Overview of potential drivers and influencing factors of market prices for regranulates.

Since, we can only process those influencing factors in the forecast model with a quantitive representation like the time series to be analyzed – in this case, it means the same granularity and a sufficiently long and up-to-date data history, we leave the interpretation of the influence of one-time and political events on recyclate prices to experts and concentrate on the quantifiable factors. Once we have identified among all available possible factors those that have predictive power, we can suitably utilize them in the model and thereby obtain more precise forecasts.

Analyzing influencing factors in combination and with time lag

We analyze the influence of external influencing factors on the target recyclate price time series using the MATCHER module of the forecasting software future. In this way, we can ensure that the influencing factors enter the forecast model with the ideal weight and the ideal direction of influence. Relationships between the influencing factors also play a role here, so we consider influencing variables not only separately but also in combination and potentially time-shifted. Although we suspected, for example, that both the prices of primary plastics and the prices of their waste could be relevant, it ultimately turns out that one of the two is dispensable for the forecast: They exhibit such a high correlation that it leads to collinearity in the model and thus brings no added value to the model. On the contrary, you get a more stable forecast model if you dispense with one of the two factors – at least in the case of additive models, which we are particularly focusing on here. Ultimately, we include the waste prices (“Waste price”) in the model with a time lag of one month (lag 1) and a positive direction of effect, as these show a slightly better performance.

Trend reversals can be anticipated through external influencing factors

In combination with the waste prices, we find in the OECD’s Business Confidence Indicator an influencing factor for which we could also demonstrate an effect in the model, namely also with a positive relation to the recyclate time series. This OECD indicator on business confidence is based on opinion surveys on the development of production, orders, and inventories of finished goods in the industry and is explicitly intended to anticipate turning points in economic activity.¹ Our analyses show that the indicator can actually achieve this in the case of recyclate prices. As with the waste prices, we can determine that the indicator has a lead time of one month on the target time series. At first glance, this is not much, but when purchasing large quantities of a raw material, it can already make a real difference.

Figure 3: Recyclate prices with the influencing factors shifted by their respective lead times (OECD Business Confidence Indicator and waste prices)

We have thus found crucial factors and carriers of relevant information for the forecast of trend reversals of the recyclate time series with the waste prices and the OECD indicator and will now look at the forecast in the next step.

The perfect settings for price forecasts

We create the forecast using the futureFORECAST module of the forecasting software future. In the module, several different forecasting models compete against each other, and among all the plausible models for the present data, the one that delivers the best results is selected using intelligent selection mechanisms. Among the tested models are various methods from the fields of statistics and machine learning, including some procedures we have developed ourselves. Thanks to the relatively long data history of the recyclate price time series, a large number of possible forecasting methods are in principle available to us here. However, we want to integrate external influencing factors (see above), so we can only use those forecasting methods that allow to integrate them and cannot only process the pure time series history. With the remaining usable forecasting methods, future supports us through intelligent mechanisms to use them in the correct way, e.g., so that model or hyperparameters are ideally chosen in the desired sense or outliers are handled in a certain way.

If we wanted, we could deal with the question of in what sense all these settings should actually be ideally made. Because we have the choice to evaluate the models using different forecast accuracy measures. In our case, we relied on a proven score from certain in-sample and out-of-sample criteria. future also allows weighting certain forecast steps more heavily than others. Here, we have particularly strongly weighted forecast step 3, i.e., the forecast for three months ahead. Ultimately, including influencing factors in our application leads to a forecast error of 1.5% – which corresponds to a reduction of the error by more than 20%!

Figure 4: Point forecasts and prediction interval at the 95% prediction confidence level for recyclate prices under the forecast model with influencing factors

Prediction intervals illustrate random fluctuations

In addition to point forecasts – a kind of best guess for future regranulate prices – we also return prediction intervals, giving us a sense of the amount of forecast uncertainty that is unexplained. This forecast uncertainty, converted into numbers, is reflected in the upper and lower limits of prediction intervals. We have chosen a prediction confidence level of 95%, the industry standard. It is clear that the prediction intervals become wider for months further in the future. The prediction intervals illustrate a behavior of the recyclate prices that cannot be explained or described by indicators or other influencing factors: The environment is quite complex; there are a multitude of influencing factors. What they are and how strong they are is subject to change as well. The fact that these are often not quantifiable is shown, for example, by the non-navigability of container ships in the Gulf of Aden mentioned by the data journalist Peter Jetzer in the Recycario report from April 2024 . By focussing on plastics and drawing on his own expertise, he is able to explain a portion of the uncertainty, while a small residual is solely due to chance. Ultimately, however, the uncertainty cannot be completely eliminated in the forecasts. But we can at least make them transparent in the form of prediction intervals.

Best-case & worst-case scenarios for a sound basis for decision-making

With the help of forecasts, we go one step further to make the uncertain future of recyclate prices a bit more tangible. Our analysis of the influencing factors of waste prices and OECD business confidence has shown a lead time of one month for both. So if we know the values for both up to and including May 2024, we can create a forecast up to June 2024 using them. That’s already something, but we’d prefer to look further into the future as we want to find our way around the new and therefore still uncertain recyclate market. These influencing factors are indeed not easy to assess, although they are at least linked to longer empirical values. For example, experts could make assumptions about their development based on current economic and political circumstances and decisions and translate his estimates into figures for possible future trends in OECD business confidence and the waste price time series. The experts could also propose several scenarios, covering optimistic and pessimistic outlooks on the development of OECD business confidence and waste prices. In studying the effects of these outlooks on the recyclate prices, we may ultimately derive a plausible range for future recyclate prices.

In our case study, we have relied on data-based scenarios instead of expert estimates: From the historical time series of OECD business confidence and waste prices, for each we have calculated a positive and a negative course to create scenario forecasts with our software. We are aware that such scenario forecasts naturally depend greatly on the assumptions made. Nevertheless, they offer a valuable insight into the sensitivity of recyclate prices with regard to the influencing factors and a possibility to incorporate important expert knowledge into the forecasts.

Figure 5: Scenario forecasts for recyclate prices for May 2024

Conclusion: The future of recyclate prices becomes more transparent through scenario forecasts

We have generated forecasts for recyclate prices for the coming months and were able to integrate decisive influencing factors that can anticipate the development to some extent. Through various realistic data-based scenarios, the future of the recyclate market has become more transparent, and purchasers can now make informed purchasing decisions based on the results.

By including influencing factors we have essentially achieved two things: 1. A higher forecast accuracy, which means that future prices can be better estimated. To return to the initial example of the gas station, this means that real money may be saved when you time refueling cleverly. While costs for gasoline are still comparatively small, they can quickly increase when large quantities of plastic are needed. The savings are accordingly high. 2. Existing practice and decades of experience in the established market could in some way be transferred to the new market or converted into quantifiable results by appropriately selecting influencing factors and forecasts.

As is often the case, more is not necessarily better: We even dispensed with an influencing factor for the sake of the model’s stability. We used the forecasting software future to generate the results. It can meet the complex requirements of a price forecast that systematically processes both the price history and other influencing factors from the economic environment and generates concrete figures for the future.

There remains an uncertainty in the forecast model, which we have depicted through prediction intervals. To explain it and interpret the results, experts from the recyclate market like Peter Jetzer are needed, whose many years of expertise enable him to, for example, relate one-time events to the recyclate market – which is something a data-based model cannot achieve. Nevertheless, through our price forecasts, the relatively new recyclate market has become significantly more transparent, and forward-looking action as well as informed purchasing behavior has become possible.

Literature Note

_{1: OECD (2024), Business confidence index (BCI) (indicator). doi: 10.1787/3092dc4f-en (Accessed on 26 June 2024)}

You would like to test future yourself right away?