The Blockchain Economy

(1)

The Block Chain Economy:

A study on the potential of implementing neural

network approaches for providing economic stability.

Author: Paul Heijen

10663533

Bachelor thesis Credits: 18 EC

Bachelor Opleiding Kunstmatige Intelligentie University of Amsterdam Faculty of Science Science Park 904 1098 XH Amsterdam Supervisor: Prof. dr. R. Valenti

Intelligent Systems Lab Amsterdam Faculty of Science

University of Amsterdam Science Park 904 1098 XH Amsterdam

(2)

Keywords:

Time series models, Forecasting, Block chain, Neural Networks, Demand and Supply, Data Volume, Transparent Digital Economy, Data continuity, LSTM, ARIMA

1 Introduction

After the second World War the economic model of capitalism has shown to be very successful in the modern world. Western economies were thriving and seemed to show potential for endless growth. However, this seemingly flaw-less model has demonstrated some cracks as was shown in earlier and recent economic depressions/crisis. The underlying cause for these economic depres-sions is a subject that has undergone much study and debate and although the exact causes might be hard to pinpoint, there are some conclusions that can be drawn from this debate about the way the economy is organized. One important conclusion is that the economy and the joint valuation of products and resources lacks transparency(Swan, 2015). This lacking transparency, gives room for manipulation of values caused by conflicting stakes for different indi-viduals, organizations, companies and even countries. The road from resources to actual products is longer than ever, and the endless number of ’middle-mans’ caused by differentiation of production makes it harder and harder to keep track of the actual value of products. However, recent developments have shown ways to tackle this lack of transparency and this will be the base of this paper(Swan, 2015).

Since the introduction of the Bitcoin in 2009 en its success in the valuta market, many scientists have studied the underlying concept of the block chain distributed database and discovered much potential for possible implications in other areas (Underwood, 2016). Its main quality that offers this great potential is the decentralized stability that it exerts (Swan, 2015). The fact that every new transaction in such a database is just incrementally adding a block, consisting of a time stamp, a transaction and a link to a previous block, to the existing block chain makes it resistant to retroactive modification and since this block chain acts as an open distributed ledger, transactions can easily be verified by the users themselves and therefore it has no need for a central trusted administrator (Nakamoto, 2008). As an autonomous transparent database the concept of block chain offers great potential for all services that require some form of registration and/or valuation done by a central administrator and in this paper the focus will be on its possible integration in the economy (Goertzel, Goertzel, and Goertzel, 2017) for reasons that were sketched in the first paragraph of this introduction. It’s peer-to-peer organizational structure could reshape the way we think about economics and the idea of decentralization and peer-to-peer is something we already see today in companies like AirBnB and uber. Moreover, the re-sulting quality of transparency should lead to an economy where every value unit can only be transferred once (which solves the double spending problem) and every transaction can be verified and studied (Saveen A. Abeyratne, 2016). In a running economy such transactions would be numerous and generate a lot of data. Big data, if you will, which is a word that triggers a proper machine learning scientist to unleash predictive algorithms on it in order to model the behavior of this data. This behavior, in terms of the economy, is the demand for certain products or resources. Predicting future demand can help balance

(4)

such an economy in terms of matching demand and supply, which in a regu-lar non predictive economy would fluctuate around a trend due to, production time as the temporal gap and resource location as the spatial gap between them (Mattila, Timo, and Jan, 2016). In an economy that is more balanced like that, overproduction and over-valuation could be a thing of the past, which would be, in a world where we getting more and more aware of the value of natural resources, a welcome result. These algorithms could be applied in all domains and scopes of the economy since everyone will have full access to the data due to the transparency of the block chain. In recent years many new types of neu-ral network structures have been developed that can deal with prediction and classification on data with a temporal aspect so a predictive algorithm could theoretically be produced for such a potential transparent economy (Jia, 2016). And since the attempt of predicting future demand is a discipline that is much broader than this proposed approach involving neural networks, for means of fair comparison, it would be really interesting to discover how the proposed approach relates to traditional methods in terms of performance. In this con-text, traditional methods would imply linear methods which are known to be limited to the modeling of just linear dependencies in the data (Claveria and Torra, 2014). Neural networks however, are proved to have the potential to discover non-linear dependencies in data and this is and interesting notion to investigate in an economical context, since economical dependencies tend to be very complex and non-linear (Vasilopoulos, 2015). However, it is known that neural network algorithm require huge amounts of training data for satisfiable prediction performance, where traditional linear methods are believed to de-liver reasonable performance on lower amounts of training data (Bontempi, Ben Taieb, and Le Borgne, 2013).

Taking the sketched baseline of a transparent digital economy as a result of block chain implementation in mind, this paper will attempt to evaluate and interpret the effect of implementing neural networks over traditional linear methods as a predicting and controlling system regulating economic processes. Interpreting economic processes in the most broad and general sense, means the focus of the proposed comparison in this paper will be embedded in a of general applicability on different types of data. Taking this focus into account, the mod-eling will be univariate, since adding multiple features would make such models product specific and in the sense of general applicability this isn’t desired for means of the described comparison.

So in summary, this paper will attempt to give a answer to the following research question:

In a transparent digital economy, can neural network algorithms help to con-trol the supply for goods by predicting their demand in such an economy in order to stabilize it and prevent the unwanted consequences of instability in terms of over valuation and over production?

(5)

The next chapter will be devoted to proposing a method that could provide an answer for the proposed research question. After that, a chapter will be used to properly present the eventual result of the proposed method. And the two consecutive chapters will embed the conclusions than can be drawn from these results altogether with critical a reflection on these conclusions.

2 Method

To give a proper answer to the proposed research question it is to be researched if creating such a system that can predict and control an economy is possible. To attempt this, a number of things will be required to actually create, simulate and evaluate some form of an economic process guided by Neural Network algo-rithms. The first one being, a database containing transactions that somehow represent economic processes. Secondly, the corpus of Neural Network algo-rithms needs to be considered to see which type of algoalgo-rithms seems most viable to deal with predictions on the eventual database. After that, the predictions need to be evaluated to see if the constructed algorithm actually has a beneficial effect on the constructed economy, in terms of defining an proper error metric as performance parameter and comparing these results with a potential bench-mark approach for demand predicting that doesn’t involve the implementation of neural networks. And finally, the actual data combined with the predictions on the test part of this data for a potentially best performing algorithm will be used as a simulation for purposed of an economical interpretation of the ob-tained results. These things will be discussed respectively in the sections of this chapter below.

2.1 Defining an economy

To piece together a simulation of an economy, the first thing to be done is to decide how such and economy would look like in terms of raw data, which will be discussed in the first subsection below. After that, it is to be decided which level/scope of economic processes is to be analyzed, which will be done in the second subsection. And finally, the last two subsections will be devoted to ex-plaining the necessity of qualifying the potential data in two different dimensions for means of being able to give a proper analyses.

2.1.1 Quantifying an economy

To build and simulate an economy, a form of data is needed that somehow represents the tension between demand and supply, which is the core of every economy. Related studies demonstrate that the most basic way to do so, is obtaining a database of historical transactions (Claveria and Torra, 2014) (Saad, Prokhorov, and Wunsch, 1998) (Cubero, 1991). These transactions inhibit a demand component as well as a supply component, since they sketch what has been sold over a certain period of time, they can be seen as the result of the

(6)

earlier proposed tension (Taylor, Menezes, and McSharry, 2006). Assuming the data represent a well functioning economy it can be said that demand and supply are already matched on a certain level (which we will try to improve). This assumption is necessary because if the goal is to predict future demand, a part of this obtained data is to play roll of historical demand so a predicting algorithm can be trained on it. So in terms of machine learning, the obtained dataset will partly function as representing the demand. This will be approximately 70 percent, as is common in machine learning, and will be used for the algorithm to train on. The remaining 30 percent of the data will then be regarded as being the actual future demand, from the point where the data was split (chronological order), and can be used to evaluate the demands predicted by the algorithm.

Ideally, such a database will have the form of a table with to columns. The first column should contain the periods of time, which ideally would be days because this can easily be extended to weekly and monthly periods while this isn’t possible the other way around. And the second column should contain the quantity of the transactions of a certain product on that period of time. This univariate approach is similar to a relate able study done by Taylor et al. in 2006 (Taylor, Menezes, and McSharry, 2006). They use the demand for electricity in the dimension of Megawatt as the variable to be predicted, which can be linearly replaced by the demand for a certain product/service in the dimension of quantities. Another study where a similar structure for the data is used is a comparative study conducted by Bajari et al. in 2015 (Bajari et al., 2015), where the demand for grocery products is to be predicted.

When visualized such a database should represent a nice demand graph which would be almost similar to a stock graph. This is desired, because in that way it is a good indicator for demand and as was shown, much related work treats data in likewise experiments this way.

2.1.2 Level of economic processes

Since it is decided how the data needs to be structured, it is now time to consider which level and scope of economic processes should be analyzed. In doing this, three things need to be considered. Firstly and most importantly, it needs to be representable in a way so that it can actually be used to answer the research question. Secondly, it has to somehow fit in the form of data we discussed in the previous section. And last but not least, the data needs to be obtainable, in terms of being freely available on the Internet since this project has no funding to acquire expensive data.

Considering the firstly mentioned requirement and what was stated in the previous section, it needs to be such that it represents a tension between ask and demand. If we add the second requirement to this, it needs to contain trans-actions on a specific category of things, which could be real products, services or a more general version of these. And to mold this to the last requirement it shows, after exploring the Internet for data whit such qualities, that most of this data is found on a single product level or multiple products within one company. This of course seems very logical since most datasets are about a single company

(7)

and very few take the effort to combine data from multiple companies. And in terms of these companies and their data, this extensive exploration shows that they mostly contain data about on product, in which they are specialized, or multiple products offered by the same company.

So in this project in terms of our research question, an economy will be con-sidered on the level of a company which offer one product or multiple products for reasons mentioned above.

2.1.3 Volume dimension of data

Looking into similar experiments regarding predicting future demand using Neu-ral Networks, it can be noticed that in most cases the quality of the eventual predictions of such algorithms very much depend on the amount of data points that are available for such an algorithm to train on (Bontempi, Ben Taieb, and Le Borgne, 2013).

And since the goal of this project is to discover if Neural Networks can ac-tually predict future demand well, in comparison to traditional methods which will be explained later in this chapter, in a broad sense, it would only be fair if we take multiple levels of data volumes into account in this research. Therefore a standard will be introduced that classifies the volume of a dataset into three different categories. The parameter(v) for this standard is calculated according to this formula:

VP roduct=

Ntransactions in dataset

Ntime periods(days)

(1) After that the value of this parameter tells us in which category the dataset of this specific product is placed as is showed by figure 1.

Figure 1: Table representing the different volume categories

In this sense, the expectation will be that Neural Networks will definitely outperform traditional demand predicting approaches when the data will be in the high volume category. But it will be really interesting to see how the results Neural Network algorithms contrast to that of traditional approaches in the lower categories.

2.1.4 Time scope dimension of data

Another interesting contrast could show itself when dealing with different time periods. Therefor, the proposed models will be tested on different time scopes,

(8)

for instance from days to weeks, to see if this shows significant differences in the predicting power of the proposed models.

This is particularly interesting in the context of this project because in terms of matching potential supply to the predicted demand, the production time of said products is key factor in this matching process. If for instance, it takes a week to produce a certain product one would be more interested in the quality of weekly predictions as opposed to that of daily predictions.

Another interesting effect of reducing daily entries to weekly entries, will be that the data will become more continuously defined, in terms of possible time lags. For instance when a product doesn’t have data for a couple of days and therefor exalts a time lag, this time lag will be consumed when the daily entries are combined to a weekly entry. The relevance of this notion will become clear when the specific qualities of different neural networks are explained which will be done in the next section.

2.2 Predicting with neural network algorithms

At this point, research has been done regarding to what kind of data needs to be analyzed to construct a satisfying answer to the proposed research question. And to elaborate on the underlying hypothesis that machine learning/neural networking can provide, more than the current traditional approaches, accurate predictions on demand in a general way, let’s shine a light on what drives this hypothesis. To do this one must be able to grasp the idea of time series fore-casting. This idea is based on the notion that the future depends on the past. In other words, there is a relation between the future and the past. However this relation tends to be very complex and hard to capture in an analytical way. Traditional statistical analyses of time series data tends to decompose the variation of data in the time series in three components being:

• Trend: The long-term change in mean level • Seasonal effect: Seasonal variation

• Irregular fluctuations: Noise. Random fluctuations

(Bontempi, 2013; Hyndman and Athanasopoulos, 2014; Kendall, 1973) This decomposition is visualized in figure 2.

In terms of this project, from an economical point of view, these first two components could prove to be very interesting, and if captured could really prove useful in terms of predicting demand. However, to capture these components with traditional linear statistical models, of which an example will be given later in this chapter, requires much knowledge about the data to optimize such a model with the right parameters. Furthermore, these models cannot interpret the third component, since it cannot be captured in linear relations. And finally, in real world problems, like our economy, the defining variables are unlikely to be linked in a linear way (Hamza¸cebi, Akay, and Kutay, 2009).

To put these problems in the context of this project, let’s backtrack a little. This project builds on the notion of a potential transparent digital economy

(9)

Figure 2: An image visualizing the decomposition of time series data, image was taken from Bontempi (2013)

and its public accessibility to the data defining it. The goal is to research the possibility of a general applicable algorithm that can provide predictions on the demand side of this economy in order to more smoothly match to the supply side of it. As was shown, traditional statistical models tend to be very task-specificly constructed and require different parameter settings for different problems, which doesn’t match with the proposed requirement of being generally applicable. Also the fact that economic data tends to be very noisy and non-linear proposes many problems for this approach.

The basis of the proposed hypothesis starts to become clear when one looks into the benefits of dealing with such problems with a neural network approach. Neural networks show the potential to deal with earlier mentioned problems, since it learns the values of said parameters itself through training steps and the neural network allows for non linear relationships to be captured in its structure and therefor has the benefit of being more flexible and generally applicable as was shown in an comparative study conducted by Swanson and white in 1997 (Swanson and White, 1997). The commonly assumed downside to this approach

(10)

would be that it requires a large rich dataset to be properly trained for viable accuracy, which in the context of an open digital economy isn’t likely to propose a problem.

Nowadays the research field of neural networking has widely expanded over the years within the discipline of Artificial Intelligence and many types are currently being used to tackle different kind of problems. The next sections will provide an overview of neural network structures that are common in predicting time series data and will shine a light on its qualities that make it so.

2.2.1 Regular Neural Network

A Neural network is an interlinked web of nodes consisting of an input layer, a hidden layer(or multiple hidden layers which is the case in a deep neural network) and an output layer. The input layer is used to receive information from the data. The hidden layer consists of nodes both connected to the output layer as well as to the input layer, and is considered to be the most important component of the network since this is the part where non-linear relations can be captured. Which is one of the qualities desired as was sketched earlier in this chapter. And the output layer is used to eventually produce a prediction ahead in time (Zhang, G 2012).

2.2.2 Feed forward Neural Network with Multilayer Perceptron A Multilayer Perceptron (MLP), is a special version of a neural network. Namely, it is a feed forward neural network, meaning that the network structure is one-directional from input to output as opposed to being cyclical, which can be the case for regular neural networks. The network is eventually trained making use of Backpropagation which feeds defined error function, resulted from comparing the output with the actual test data, back into the network to get it properly trained by adjusting the weights of the connections between the nodes in the network. The benefit of this network structure is that multiple time points can be consumed by the input layer while remaining its temporal meaning. This quality excerts itself in being able to receive lagged observations as input in the form of (Xt, Xt−1, ..., Xt−p). Such a feedforward neural network structure

is visualized in figure 3.

Such a network can also be described by the following equation:

Xt= f (β0+ q X j=1 βjg(xt−1φij+ φ0j)) (2) { φij = 0, 1, ..., p, j = 1, ..., q } { βj= 0, 1, ..., q }

where f is the output function, g is the activation function, p is the number of inputs, q is the number of neurons in the hidden layer, xtis the output, xt−1

(11)

Figure 3: A feedforward Neural Network with three output layers

φij are the weights connecting the hidden layer with the input (Claveria and

Torra, 2014).

2.2.3 Neural Network with Long Short Term Memory nodes

A Long Short Term Memory (LSTM) neural network is a different type neural network, namely a recurrent neural network(RNN). The quality RNN’s exhibit that makes this type of neural network interesting for a time series forecasting approach is that it can take (remember) information from earlier data points to be used in its prediction, by it having cyclic feedback connections between its nodes. This can partly be done with the earlier mentioned MLP by giv-ing it multiple input nodes lagged in time, but in that case the length of the memory would be non arbitrary because its predefined in the network struc-ture. RNN’s as told, do posses the ability to remain information from data points earlier processed, however these long-term dependencies aren’t learned by the network itself, as was shown by a study that investigates the ability of

(12)

long-term dependencies being captured by regular RNN’s (Bengio, Simard, and Frasconi, 1994). However, they are built in the structure and therefor require some form of knowledge about these dependencies to construct such a network. And since this doesn’t match with the being generally applicable requirement of this project, a special version of RNN needs to be considered that is actu-ally able to learn these dependencies. And that will be the LSTM version of an RNN, designed in reaction to problem of RNN’s sketched above. The main difference with regular RNN’s is that it contains a recurrent node, an LSTM node/cell, that is able to store information for long or short periods of time and doesn’t make use of an activation function, as opposed to regular RNN’s. This way, the stored value will not be iteratively washed away over time as the model is trained. Such an LSTM node consist of three gates being an input, forget and output gate which is visualized in the figure 4.

Figure 4: Diagram of the structure of a single LSTM node/cell. Image taken from Jia (2016)

(13)

with a certain value. These weights are optimized during the training of the network and therefor able to catch long term dependencies in terms of them having a possible beneficial effect on the accuracy of the predictions made by the network (Jia, 2016). This quality could prove an LSTM neural network approach is more suitable for data that might contain interdependencies in data points that are far apart in time. Adding to this that it’s able to handle time series data with time lags of unknown length makes it even more interesting. In terms of this project, this is a desirable quality because most datasets that match the requirements proposed in the previous section do not contain data points for days where no sales where made, and therefor might contain many of those time lags.

2.3 Suggested interpretation and evaluation of results

Since the ground rules and structure of the proposed framework has now been explained, this section will be devoted to laying out the procedures that will be used to proper evaluate the eventual results. This will be done by firstly providing a proper metric that indicates the performance of said algorithms. And finally, a traditional forecasting approach, thats doesn’t involve neural net-works, will be presented to function as a benchmark for evaluating the proposed neural network approaches.

2.3.1 Evaluation metric

To evaluate the predicting accuracy of the constructed models, the error metric of Rooted Mean Squared Error(RMSE) will be used. This is the most commonly used error metric to evaluate time series predicting algorithms in the field of Machine Learning and is also applicable to linear models. As the name implies, it’s an indicator of the average error on all data points in the test set. To relate this to the goals of this project, where we are trying to predict demand in order to match supply to it, it’s value is a nice generally understandable value to show the quality of a predicting model, since it’s in the same dimension as the value to be predicted, which in this project will be product quantities.

The RMSE is calculated according to the following equation:

RM SE = n X t=1 (Ypred− Yactual)2 n (3)

Where n is the number of data points in the test set, Ypred is the value of a

certain point and Yactual is the actual value of the point that was predicted.

2.3.2 Benchmark algorithm

To properly evaluate the proposed non-linear approach of predicting with neural networks, a decent linear approach is needed as a benchmark. The literature shows that the most widely used linear approach for time series forecasting

(14)

is Autoregressive Integrated Moving Average(ARIMA) (Hamza¸cebi, Akay, and Kutay, 2009). Related work has even shown that this approach, when properly parametrized on the data, can outperform neural networks (Claveria and Torra, 2014). But it will be interesting to see how it performs when more generally applied and refined to the same extent as will be done for the neural network approaches. xλ_τ = Θs(L s_)θ(L) Φ(Ls_)φ(L)∆D s∆d t (4)

where Θs(Ls) = (1 − ΘsLs− Θ2sL2s− ... − ΘQsLQs) represents the seasonal

moving average polynomial, Φs(Ls) = (1−ΦsLs−Φ2sL2s−...−ΦP sLP s) defines

the seasonal autoregressive polynomial, θ(L) = (1 − θ1L1− θ2L2− ... − θqLq) is

a regular moving average polynomial, φ(L) = (1 − φ1L1− φ2L2− ... − φpLp) is a

regular autoregressive polynomial, λ is the transformation value, ∆D

s represents

the seasonal difference operator, ∆d _{is the regular difference operator, S is the}

time period length and t is the innovation which can be regarded as behaving

like white noise.

2.4 Economical interpretation and simulation

To actually see what effect the implementation of a Neural Network Forecast model has as opposed to the implementation of the traditional ARIMA model, the resulting predictions shall be translated to corresponding differences in sup-ply and demand. For this analyses, the point where the data is split into train data(first 70 percent) and test data(last 30 percent) shall be viewed as the start-ing point for the economical simulation. The test(real) data will be assumed to resemble the actual demand of this product and the predicted data shall be viewed upon as the actual supply of this product. This assumption is fair, since a company will produce its products based on what they think they can actually sell. The models provide these predictions and the eventual difference with the actual demand data will resemble the mismatch of supply and demand. The goal of this project can be viewed upon as minimizing this mismatch in terms of product quantities, it would be really interesting to visualize and quantify this effect.

To proper visualize the effect of the implementation of different models a graph is produced, where the y-axis resembles the difference between supply and demand (prediction minus actual) and the x-axis shows the steps in time. The mismatch of supply and demand is also quantified in an absolute numbers by subtracting the prediction data with the actual data and summing the nega-tive values to a number that resembles the quantity of underproduction, which results in over-valuation (Samuelson, 1964), and summing the positive values to the quantity of overproduction. These numbers can also be interpreted as representing the area between the difference graph and the perfect match graph, which would be a straight line along the x-axis.

(15)

3 Results

As the methodology of this project was explained in the previous chapter, this chapter is reserved for presenting the actual results of the proposed method. These results will be presented in a similar section structure as was used in the method chapter.

3.1 Analyses of data acquired

With the proposed data requirements from the previous chapter in mind, two datasets were eventually acquired that are able to show the contrast between the neural network approach and the linear approach(ARIMA).

The first dataset, being the Peerby dataset, contains over 310 different prod-uct types and has a total amount of 452.515 transactions. These transactions are spread over a period of 491 days and the average number of transactions per product is 1455. This results in some products having very few transactions and some products having relatively numerous transactions. So in terms of the ear-lier mentioned volume qualification of the data, the dataset can be considered to contain low volume data and medium volume data.

The one thing left to consider here, is high voluminous data, which was eventually found making use of the Google Big Query API. From this API a dataset was downloaded that contained the number of booked taxi rides per day for the Yellow Cab taxi company in New York City. Taking into account that New York City is one of the most crowded cities on earth, one can imagine this dataset to contain numerous transactions as you can see in the figure 5.

Figure 5: Analyses of acquired data

So data about 311+1 products is acquired, which contain an example of high volume date, medium volume data and high volume data. These datasets contain daily transactions which can easily be extended to weekly transactions

(16)

by simply summing up the daily quantities over a week. In the next sections, the results of the algorithms will be shown and explained.

3.2 Training the models

In this section, it will be demonstrated how the results per algorithm were ob-tained. This will be done by demonstrating how the eventual output of the algorithms to be compared for certain products will look like. Since these algo-rithms will be trained for more than 300 products it would be quite redundant to show this for each product individually but in the section after this an overview will be given of all the results.

(a) Regular NN (b) MLP

(c) LSTM (d) ARIMA (benchmark)

Figure 6: Resulting graphs of DAILY analyses of the product ’Bakfiets’

As an example, the modeling of the dataset of the Peerby product: ’Bakfiets’ will be demonstrated. The data of this product classifies as Medium Volume dataset since the calculation of the introduced parameter v shows us it that in falls within this category as demonstrated below:

VBakf iets=

Ntransactions in dataset

Ntime periods(days)

=6085

(17)

So since 2 < 12.4 < 20, it can be classified as having a medium Volume. The current framework of algorithms is configured in such a way that for ev-ery product, it will produce a value for the RMSE and a graph containing the ac-tual quantity values over time, the model fit on the training data, and the result-ing fit of the test data. Per product this will be done for the data containresult-ing the original daily time period as well as for the same data reduced to weekly quan-tities. So in total, each product will produce Nalgorithms∗ Ndif f erent time periods

(which is 4 ∗ 2 = 8 in the context of this project) different graphs as well as values for RMSE. For the ’Bakfiets’ example theses results are shown in figure 6 for the daily segment and in figure 7 for the weekly segment.

(a) Regular NN (b) MLP

(c) LSTM (d) ARIMA (benchmark)

Figure 7: Resulting graphs of Weekly analyses of the product ’Bakfiets’

As you can see in the figures shown above, on a daily basis, all the algorithm perform quite well regarding the prediction of future demand of the product Bakfiets’. So it can be concluded that the example given above shows very promising results. However for a proper analyses of the qualities of each al-gorithm, the procedure, as demonstrated above, should be conducted for all products spread over different volume categories which will be done in the next section.

(18)

3.3 Interpretation and evaluation of results

Now a demonstration of the procedure for one product has been given this section will attempt to show all the results produced by the different models. The next three, one for every volume category, subsections will each show the test RMSE for each algorithm per product in that specific volume category.

And important thing to notice at this point is that for the medium and low volume products not every algorithm was able to produce a result. In terms of fair comparison, only the products that produced a model for each individual algorithm will be taken into account.

3.3.1 Low volume

(19)

The results of the low volume products are shown in figure 8. The first column shows the name of each individual product and the second column shows the number of transactions the data of that product actually contains. The table is sorted on the value of the transactions column so in can be more smoothly compared to the tables of the other volume categories.

The second(daily) and third(weekly) section of the table show us the result-ing RMSE values per algorithm plus a column that is reserved for showresult-ing the best performing algorithm per product. The lowest RMSE value per row per time period is marked green to proper visualize the performance contrast among the different products.

A first glance at this result table on the daily side, indicates that in this volume category there isn’t a truly conclusive winner to point out. This becomes even harder when the weekly part is also taken into account. Nonetheless, this table shows us that in daily section, in general the LSTM algorithm seems to demonstrate the best performance and in the weekly section, the LSTM as well as the MLP algorithm compete in being the best performers. In table 1, shown below, the wins per algorithm are summed so we can conclude that in both the weekly and the daily section that the LSTM algorithm is the winner. This seems fair because it can definitely be stated that this algorithm overall outperforms the benchmark ARIMA model. So taking the proposed research question of this project into account, this already demonstrates quite promising results when the ARIMA model performance is taken as a benchmark.

However, it might be much more interesting to see how the algorithms per-form on the higher volume categories because these are more likely to appear in the context of a transparent digital economy.

Table 1: Performance analyses in low volume data (18 products)

(a) Daily Algorithm Nwins LSTM 10 Regular NN 6 Arima 2 MLP 0 (a) Weekly Algorithm Nwins LSTM 7 MLP 7 Regular NN 2 Arima 2 3.3.2 Medium volume

For the medium volume data a similar table was produced, displayed in figure 9 on the next page. This table shows more conclusive results in the daily segment in comparison to the low volume products. But in the weekly segment it shows a similar diffuse pattern of winning algorithms. Taking a look at the win tables (table 4 ), once again the LSTM algorithm proves to be the best performing algorithm in both segments, although much more decisive in the daily segment.

(20)

(21)

Table 4: Performance analyses in medium volume data (31 products) (a) Daily Algorithm Nwins LSTM 23 Regular NN 5 MLP 3 Arima 0 (a) Weekly Algorithm Nwins LSTM 13 Arima 8 MLP 4 Regular NN 6 3.3.3 High volume

As the result table in figure 10 shows, in this volume category the MLP approach proves to be the winner. It proves to be a more decisive winner in the daily segment as opposed to the weekly segment and the possible reasons for this result shall be evaluated in the next subsection.

Figure 10: High Volume: RMSE values per algorithm

3.3.4 Evaluating results in theoretical context

To get a proper feel of why the algorithms produced the presented results, this section will attempt to give an explanation for these results based on the theoretical qualities per algorithm that were described in the previous chapter. Low Volume data The result table shows scattered results in the low volume category. In the daily segment, a careful conclusion can be made that the LSTM algorithm shows the best performance. From a theoretical point of view, this could be expected since these low volume products contain many time lags, in terms of containing days without transactions. And in the previous chapter, it was explained that the LSTM approach is specifically structured to deal with such time lags. This notion becomes even more clear when we take the weekly segment into account where LSTM doesn’t appear to be a decisive winner. By reducing the daily entries to weekly entries many daily time lags are consumed and therefor the data automatically becomes more continuously defined, and the advantages of the LSTM approach become less appear-ant. The occasional win of the ARIMA model in this volume category can be explained by the fact

(22)

that ARIMA is supposed to be better equipped to deal with training data of smaller sizes.

Medium Volume data Similar statements can be made about the perfor-mance of the LSTM algorithm in medium volume category since a similar pat-tern is visible. In this category the daily segment doesn’t show any wins for the ARIMA model which could be expected since these products have data with more volume, which shifts the balance of power more towards the machine learning approach. The relatively many wins for the ARIMA approach in the weekly segment can be explained by the fact that again the data becomes more continuously defined but hasn’t yet reached a proper level of volume to give the neural networking approaches the learning advantage.

High Volume data Unfortunately, for this volume category there is only one product available so conclusions about this category are not as well funded as for the lower categories. However, the notion that the neural network approach with an MLP classifier shows the best result in this category can be explained in the theoretical context of the previous chapter. The fact that the LSTM algorithm shows a significantly worse performance than in the lower volume categories could partly be explained by the notion that the NYC taxi dataset doesn’t contain any time lags which gave this approach the advantage in the lower volume categories. Also it can be stated that in interpreting this data, there doesn’t seem to be any relevant long-term dependencies. Short-term de-pendencies seem much more relevant if one takes into account that the number of taxi rides are significantly correlated with the specific day of the week. And since the MLP is especially structured to deal with short-term dependencies it seems plausible that it proves to be the winner for this product. Moreover, it can be concluded that neural networking approaches prove to be the better per-formers when exposed to a huge amount of training data. This is proved, by the relatively bigger difference in the resulting RMSE values between the ARIMA approach and the neural networking approaches as opposed to the lower volume categories.

3.4 Economical interpretation and simulation

In this section, an attempt will be made to put the results in an economic interpretation by using the prediction as a simulation as was proposed in the previous chapter. Again this would be quite redundant and page-filling to do this analyses for each product so for means of demonstration an example will be given for a certain product where the LSTM algorithm outperforms the other models in both the daily and the weekly segment. The example shown in figure 11 below conducts this economical simulation for the product: ’schuurmachine’, and its quantification in economic values is displayed in figure 12.

The results shown above prove that in the situation where the LSTM is the best predicting algorithm (which was shown to be the case for many products),

(23)

(a) LSTM Daily (b) ARIMA Daily

(c) LSTM Weekly (d) ARIMA Weekly

Figure 11: Economic simulation where the x-asis represents the optimal situation of perfectly matched supply and demand, and the simulation graph represents the actual difference of this match and supply. The area colored green can be interpreted as the quantification of overproduction and the area in red as the quantification of under production (over valuation).

using this algorithm to regulate the production does indeed have a significantly beneficial effect as opposed to a situation where a traditional linear model is applied.

(a) Daily quantification (b) Weekly quantification

(24)

4 Conclusion

This project grew root in the basic idea that the current way our economy is arranged, is a way that has grown old and new technologies and environmental incentives could pave the way for a new economic model. As was shown in the first chapter, many recent studies show that the block chain technology proved itself to be able to deal with much of the problems embedded in the current model. Its potential qualities of transparency and trustworthiness, as a result of autonomous self-regulation, can no longer be denied by the economic world. And the recent emergence and success of block chain platforms like Ethereum prove that the balance is indeed shifting in this direction.

Assuming such a potential economic model, the focus of this project was aimed on the possible implementation of neural networks in such an economy, with the purpose of controlling its demand component by predicting its demand component. To research this hypothesis, multiple datasets of different products were acquired to see if neural networks could indeed, in a general applicable way, outperform traditional linear modeling methods in a time series forecasting per-spective. Taking into account that product data in such a potential block chain regulated economy would involve multiple levels of volume and continuity in terms of the amount and spread of data points, this project focused on analyz-ing neural network, in contrast with traditional linear methods, performances on this different levels. This was done by categorizing the product datasets by these levels and analyzing their prediction errors. And to put this all in an economical context, these predictions were finally used to perform an economic simulation to quantify these results in economically interesting values.

The results show, that neural network approaches can indeed outperform traditional linear methods when generally applied. As was shown, the proposed low volume category showed that neural network approaches can outperform the benchmarked ARIMA approach even when exposed to a low amount of data. It also demonstrated that an LSTM approach really proves to exalt when data is not continuously defined. This conclusion became even stronger when the higher volume categories were considered were. In the medium volume category, the LSTM was a real decisive winner in the daily segment, as expected. The high volume category showed that the LSTM approach lost it advantage here since the data had an optimal continuity, but still all the Neural Network strategies proved their qualities when exposed to a huge amount of data, by outperform-ing the benchmarked ARIMA approach. The final example of the economical interpretation showed us that this difference in prediction performance can re-sult in a significantly better situation in terms of matching supply to demand, quantified in values for overproduction and underproduction (leading to over valuation).

In a transparent digital economy, can neural network algorithms help to con-trol the supply for goods by predicting their demand in such an economy in order to stabilize it and prevent the unwanted consequences of instability in terms of

(25)

over valuation and over production?

So in summary, by taking the initially proposed research question into ac-count, it can be concluded that the possible implementation of neural network algorithms in such a transparent digital economy, facilitated by block chain, does indeed show huge potential in terms of predicting and controlling the supply side of such and economy as opposed to to tradition linear forecasting approaches. Time series forecasting with Neural network approaches definitely show their theoretical advantages when generally applied to different levels of data volume and continuity. One of the most interesting parts of this conclusion would be that this statements even holds for low volume datasets, although less conclu-sive when compared to higher volumes. But in the context of full transparency, provided by the block chain, acquiring enough data isn’t likely to propose a real problem.

5 Discussion

To put these conclusions in a broader scientific context, this chapter will be devoted to, first, giving a critical evaluation of the results found by giving sug-gestions to add more weight to the made conclusions and finally will offer some interesting directions of research to address in future work.

Since conclusion were made about different levels of data volume and conti-nuity, it would be really interesting to further expand this analyses by adding many more data sets that qualify differently in these dimensions. This would serve two purposes, namely it would provide more weight to the made conclu-sions, since a limited number of datasets was used, and furthermore it could offer more detailed insight because by adding more datasets, the said dimen-sions can be further differentiated and therefor offer more specific insight. For instance, if one is able to acquire data for more than 10.000 products, the vol-ume classification could be expanded to more than just 3 different categories and therefor offer a more differentiated insight. For the continuity dimension, this further differentiation could be offered by using data that extend over a longer period of time so that its data points can undergo multiple levels of re-duction without shrinking the database too much. Since the data for the Peerby products only extended over 491 days, only one level of reduction could be made without shrinking the data to an analytically impossible level, namely from days to weeks. Longer datasets, could be reduced from days, to weeks, to months to even years if long enough. In terms of this project, this would be especially interesting for the high volume category since this project only considered one product in this category.

Another point of discussion would be the fact that this project considered only the neural network approaches from the broad corpus of Machine Learning algorithms. This focus was chosen because this type of algorithms have the most promising theoretical qualities and are most commonly used, and therefor more advanced, in the context of time series forecasting. However recent studies

(26)

show the emergence of other machine learning algorithm types as significant contenders in the discipline of time series forecasting. Unfortunately, it was beyond the scope of this project to consider these algorithms as well so this would propose another interesting direction for future research.

5.1 Future work

If we transcend from the context of this project in a more broad perspective, other interesting directions for future research become clear. As the research question of this project implied, this project aimed to control the supply compo-nent of an economy by predicting the supply compocompo-nent. However, it might be equally interesting to see if the demand component of such an economy could be controlled as well. Since, demand and supply are the most significant variables in determining the price of a certain product, it could prove quite interesting to study the effect of price alterations on the demand function. This focus on the ’other side of the dime’ of such an economy could very well result, when com-bined with focus of this project, in a perfectly stable economy where supply and demand are perfectly matched. As a result, over production and over valuation could really become a thing of the past and therefor resources will be optimally utilized. Which in a world where we becoming more and more aware of the value and scarcity of said resourced, would be a very, very welcome result.

6 Acknowledgements

The contributors of this project would like to thank the peer-to-peer sharing platform Peerby for providing an interesting dataset set that proved to match perfectly with the focus of this project.

7 References

Bajari, Patrick et al. (2015). “Machine Learning Methods for Demand Estima-tion”. In: American Economic Review 105.5, pp. 481–85. doi: 10 . 1257 / aer.p20151021. url: http://www.aeaweb.org/articles?id=10.1257/ aer.p20151021.

Bengio, Y., P. Simard, and P. Frasconi (1994). “Learning long-term depen-dencies with gradient descent is difficult”. eng. In: Neural Networks, IEEE Transactions on 5.2, pp. 157–166. issn: 1045-9227.

Bontempi, Gianluca (2013). Machine Learning Strategies for Time Series Pre-diction. University Lecture.

(27)

Bontempi, Gianluca, Souhaib Ben Taieb, and Yann-A¨el Le Borgne (2013). “Ma-chine Learning Strategies for Time Series Forecasting”. In: Business Intel-ligence: Second European Summer School, eBISS 2012, Brussels, Belgium, July 15-21, 2012, Tutorial Lectures. Ed. by Marie-Aude Aufaure and Es-teban Zim´anyi. Berlin, Heidelberg: Springer Berlin Heidelberg, pp. 62–77. isbn: 978-3-642-36318-4. doi: 10 . 1007 / 978 - 3 - 642 - 36318 - 4 _ 3. url: http://dx.doi.org/10.1007/978-3-642-36318-4_3.

Claveria, Oscar and Salvador Torra (2014). “Forecasting tourism demand to Catalonia: Neural networks vs. time series models”. In: Economic Modelling 36, pp. 220 –228. issn: 0264-9993. doi: http://dx.doi.org/10.1016/j. econmod.2013.09.024. url: http://www.sciencedirect.com/science/ article/pii/S0264999313003842.

Cubero, Robert Gri˜n´o (1991). “Neural networks for water demand time se-ries forecasting”. In: Artificial Neural Networks: International Workshop IWANN ’91 Granada, Spain, September 17–19, 1991 Proceedings. Ed. by Alberto Prieto. Berlin, Heidelberg: Springer Berlin Heidelberg, pp. 453–460. isbn: 978-3-540-38460-1. doi: 10.1007/BFb0035927. url: http://dx.doi. org/10.1007/BFb0035927.

Goertzel, Ben, Ted Goertzel, and Zarathustra Goertzel (2017). “The global brain and the emerging economy of abundance: Mutualism, open collabo-ration, exchange networks and the automated commons”. In: Technologi-cal Forecasting and Social Change 114, pp. 65 –73. issn: 0040-1625. doi: http : / / doi . org / 10 . 1016 / j . techfore . 2016 . 03 . 022. url: http : //www.sciencedirect.com/science/article/pii/S0040162516300117. Hamza¸cebi, Co¸skun, Diyar Akay, and Fevzi Kutay (2009). “Comparison of direct

and iterative artificial neural network forecast approaches in multi-periodic time series forecasting”. In: Expert Systems with Applications 36.2, pp. 3839 –3844. issn: 0957-4174. doi: http://dx.doi.org/10.1016/j.eswa.2008. 02.042. url: http://www.sciencedirect.com/science/article/pii/ S0957417408001383.

Hyndman, Rob J. and George Athanasopoulos (2014). “Chapter 6: Time series components”. English. In: Print edition. OTexts.com [Heathmont?, Victoria], 291 pages ; isbn: 9780987507105.

Jia, Hengjian (2016). “Investigation Into The Effectiveness Of Long Short Term Memory Networks For Stock Price Prediction”. In: CoRR abs/1603.07893. url: http://arxiv.org/abs/1603.07893.

Kendall, Maurice George (1973). Time-series / [by] M.G. Kendall. eng. London: Griffin. isbn: 0852642202.

Mattila, Juri, Seppa Timo, and Holmstram Jan (2016). Product-centric Informa-tion Management: A Case Study of a Shared Platform with Blockchain Tech-nology. UCAIS Berkeley Roundtable on the International Economy, Working

(28)

Paper Series. UCAIS Berkeley Roundtable on the International Economy, UC Berkeley. url: http://EconPapers.repec.org/RePEc:cdl:ucbrie: qt65s5s4b2.

Nakamoto, Satoshi (2008). Bitcoin: A peer-to-peer electronic cash system,” http://bitcoin.org/bitcoin.pdf. Saad, E. W., D. V. Prokhorov, and D. C. Wunsch (1998). “Comparative study

of stock trend prediction using time delay, recurrent and probabilistic neural networks”. In: IEEE Transactions on Neural Networks 9.6, pp. 1456–1470. issn: 1045-9227. doi: 10.1109/72.728395.

Samuelson, Paul Anthony (1964). “Theoretical notes on trade problems”. In: The review of economics and statistics 46.2, pp. 145–154. issn: 00346535. Saveen A. Abeyratne, Radmehr P. Monfared (2016). “Blockchain ready

manu-facturing supply chain using distributed ledger.” In: issn: 2321-7308. url: https://dspace.lboro.ac.uk/2134/22625.

Swan, Melanie (2015). Blockchain : Blueprint for a New Economy. Vol. First edi-tion. O’Reilly Media. isbn: 9781491920497. url: http://search.ebscohost. com.proxy.uba.uva.nl:2048/login.aspx?direct=true&db=nlebk&AN= 943117&site=ehost-live.

Swanson, Norman R. and Halbert White (1997). “Forecasting economic time series using flexible versus fixed specification and linear versus nonlinear econometric models”. In: International Journal of Forecasting 13.4, pp. 439 –461. issn: 0169-2070. doi: http://dx.doi.org/10.1016/S0169-2070(97) 00030-7. url: http://www.sciencedirect.com/science/article/pii/ S0169207097000307.

Taylor, James W., Lilian M. de Menezes, and Patrick E. McSharry (2006). “A comparison of univariate methods for forecasting electricity demand up to a day ahead”. In: International Journal of Forecasting 22.1, pp. 1 –16. issn: 0169-2070. doi: http : / / dx . doi . org / 10 . 1016 / j . ijforecast . 2005 . 06.006. url: http://www.sciencedirect.com/science/article/pii/ S0169207005000907.

Underwood, Sarah (2016). “Blockchain Beyond Bitcoin”. In: Commun. ACM 59.11, pp. 15–17. issn: 0001-0782. doi: 10.1145/2994581. url: http:// doi.acm.org.proxy.uba.uva.nl:2048/10.1145/2994581.

Vasilopoulos, Athanasios (2015). “Linear and Non-Linear Regression: Power-ful and Very Important Forecasting Methods”. eng. In: Expert Journal of Business and Management 3.2, pp. 205–228. issn: 2344-6781. url: https: //doaj.org/article/e5a0936ce3e44ae0a8b153fc6da5ba8c.

Zhang, G (2012). eng. url: http://search.credoreference.com/content/ entry/sprnatural/neural_networks_for_time_series_forecasting/0.

The Blockchain Economy