Challenges of the 2020 season

Organizer: Rudy Morel, Web Developer: Tanguy Marchand, Beta-Testers: Tomás Angles, Florentin Guth, Tanguy Marchand, Rudy Morel, Gaspar Rochette, Louis Thiry, John Zarka, Supervision: Stéphane Mallat

You can find the list of the winners at the end of 2020 of these challenges here. The challenges continue to run afterwards, and the leaderboards are still updated, but the list of the preceding year winners is fixed.

Where will the next trade take place?

Given recent trades and order books from a set of trading venues, predict on which trading venue the next trade will be executed.

Community forum for sharing ideas and making faster progress:

Additional information can also be found on this forum and after registering on the Challenge Data website.

Predicting lung cancer survival time
The challenge proposed by Owkin is a supervised survival prediction problem: predict the survival time of a patient (remaining days to live) from one three-dimensional CT scan (grayscale image) and a set of pre-extracted quantitative imaging features, as well as clinical data.
Predicting response times of the Paris Fire Brigade vehicles
Your task will be to predict the delay between the selection of a rescue vehicle (the time when a rescue team is warned) and the time when it arrives at the scene of the rescue request (manual reporting via portable radio).

A first solution from a challenger has been published on GitHub:
It was not possible for us to wait any longer for the publication of this fabulous work. We will as soon as possible enhance it with the contributions of other challengers who also agreed to share their work with us. Thanks to all of them very much for the work carried out and shared.

What is extraordinary with this type of challenge is that the multitude of minds leads to the design of a multitude of methods which when combined lead us to a amazing solution that woud have not been possible to reach with the involvement of only one mind.
Asset production estimation

The goal of this problem is to estimate the production of a group of industrial assets, based on daily measurements and capacity constraints.

Considering a number of industrial installations AiA_i with nominal production capacities CiC_i , JJ daily measurements (J=5J=5 ) (xi,w,ljx^j_{i,w,l} ) are carried on each asset ii , for the week ww , for the weekday ll (l=1...7l=1...7 ), for the measure type jj , in order to detect patterns in operations that affect production.

The assets AiA_i are gathered into KK disjoint groups (K=2K=2 ). For each group, installations report the actual production levels yk,wy_{k,w} at an aggregated level, on a weekly basis.

For each group kk , the goal is to predict y^k,w\hat{y}_{k,w} , as the sum of the productions of all assets in that group (y^k,w=iC^i,w\hat{y}_{k,w} = \sum_i \hat{C}_{i,w} ) under the constraint that an asset production is smaller than its maximum capacity for each week, i.e.

0<=C^i,w<=Cimaxw 0 <= \hat{C}_{i,w} <= C^{max}_i \quad \forall w

C^i0,w0\hat{C}_{i_0,w_0} is to be estimated as a function of the measures pertaining to asset i0i_0 and week w0w_0 :

C^i0,w0=f(xi0,w0,lj)j1,...J,l1,...,7 \hat{C}_{i_0,w_0} = f(x^j_{i_0,w_0,l})\quad \forall j \in 1, ... J, \quad \forall l \in 1, ..., 7


  • The metrics can only be used to assess the production of the asset where these were taken

  • An asset usually cannot produce more than its nominal capacity, but sometimes spikes in production can go up 120% the nominal capacity.

  • The target is reported weekly while the measures are daily. Hence, the measures corresponding to the target for week ww are the measures of all the days in week ww .

  • Other factors might impact productivity, like economic conditions (i.e. demand-linked curtailment), but as a first approach, we assume such effects are not significant.

  • Some series might be lacking for some assets and are filled entirely with None values

  • These measurements may correspond to incidents or maintenance, which impact the productivity of the assets, possibly with varying significance (e.g. large v.s small incidents, …).

PhotoRoom Object Segmentation from Synthetic Images
Provided with an image of an object, the goal is to segment the salient (main) object in the scene. For each pixel of the input image, the model must categorize it as foreground or background.
Prediction of direction of Bitcoin based on sentiment data
The problem is a classification challenge that aims at building investment strategies on cryptocurrencies based on sentiment extracted from news and social networks. For each trading hour we have counted the occurence of some terms like for example "adoption" or "hack" in a selected number of influential twitter accounts as in some forums like Bitcointalk. We have created 10 different themes, some positives and others negatives and we summed the counts of the words corresponding, before normalising them. For a given sample and a given theme we use the counts of each of the 48 last hours, we Z scored these counts, and we multiplied the result by the average hourly count during the period divided by the average hourly count during all the training period. For a theme T in timestamp i, with lag k ($`k\in[\![0;47]\!]`$) the value F of the feature will be: ```math F_{i,k}=\frac{T_{i,k}-\overline{T_{i}}}{\sqrt{\frac{1}{47}\sum\limits_{j=0}^{47}{(T_{i,j}-\overline{T_{i}})^{2}}}}*\frac{\overline{T_i}}{\overline{T}} ``` We added 5 features corresponding to the price return of the last hour, the last 6 hours, the last 12 hours, the last 24 hours and the last 48 hours normalised by the volatility during the 48 hours. The aim is to predict if the return of Bitcoin in the next hour will be more than 0.2%, between -0.2% and 0.2%, or less than -0.2%. The 0.2% level is the 66.7% percentile of the distribution. The metric used for this problem is the logistic loss, defined as the negative log-likelyhood of the true labels given the classifier's predictions. The true labels are encoded as a 1-of-3 binary indicator matrix Y, ie $`y_{i,k}=1'`$ if sample i has label k taken from a set of 3 labels ( less than -0.2%, between -0.2% and +0.2%, more than 0.2%). For P a matrix of probability estimates with $`p_{i,k}=Pr(t_{i,k}=1)'`$, the log loss function is defined as ```math L_{log}(Y,P)=-log{Pr(Y|P)}=-\frac{1}{N} \sum_{i=1}^{N} \sum_{k=1}^3{y_{i,k}log(p_{i,k})} ``` The lower the score, the better.
Deep Hedging for an Equinoxe
The goal is to determine 3 optimal strategies to hedge dynamically an exotic option on a complex market
Dyni Odontocete Click Classification, 10 species [ DOCC10 ]
The goal is to classify each click according to the corresponding emitting species. The 10 species are : (0) Gg: Grampus griseus- Risso's dolphin (1) Gma: Globicephala macrorhynchus- Short-finned pilot whale (2) La: Lagenorhynchus acutus- Atlantic white-sided dolphin (3) Mb: Mesoplodon bidens- Sowerby's beaked whale (4) Me: Mesoplodon europaeus- Gervais beaked whale (5) Pm: Physeter macrocephalus - Sperm whale (6) Ssp: Stenella sp.Stenellid dolphin (7) UDA: Delphinid type A - a group of dolphins (species not yet determined) (8) UDB: Delphinid type B - another group of dolphins (species not yet determined) (9) Zc: Ziphius cavirostris - Cuvier's beaked whale The metric is the accuracy.
AI for Meter Reading

The goal of this challenge is to design an algorithm reading the consumption index from a valid picture of a meter.

Predict sex from brain rhythms

In this dataset, we try to predict the gender of someone based on 40 windows of 2 seconds taken during sleep.

Metamodels to improve energy consumptions and comfort control in big buildings
This challenge aims at introducing a new statistical model to predict and analyze energy consumptions and temperatures in a big building using observations stored in the Oze-Energies database. Physics-based approaches to build an energy/temperature simulation tool in order to model complex building behaviors are widespread in the most complex situations. The main drawback of using highly sophisticated softwares such as TRNsys or EnergyPlus to simulate the behavior of transient systems is the significant computational time required to train such models, as they integrate many noisy sources and a huge number of parameters, and require essentially massive thermodynamics computations. The most common approach is usually either to simplify greatly the model using a schematic understanding of the system, or to run complex time- and resource-consuming campaigns of measurements where trained professionals set the parameters characterizing the physical properties of the system. Even after such campaigns, calibrating these models based on real data obtained from numerous sensors is very difficult. Energy models of buildings depend on a certain number of parameters which influence the accuracy of the simulation. In order to analyze and predict future energy consumptions and future temperatures of a building, it is first necessary to calibrate the model, i.e. find the best parameters to use in the simulation tool so that the model produces similar energy consumptions as the data collected. This usually requires thousands of time-consuming simulations which is not realistic from an industrial point of view. In this data challenge, we propose to build a simplified metamodel to mimic the outputs of the physical model, based on a huge number of stored simulations. The weather conditions are obtained hourly from real sensors to store the real information relative to noisy sollicitations of the buildings. The building management system (cooling scheduling, heating scheduling, ventilation scheduling) is chosen by energy managers to provide realistic situations to the physical model. The other unknow parameters charaterizing the behavior of the building (air infiltration, capacitance, etc.) are chosen in a grid by energy managers to describe all realistic situations. For each situation specified by these input data, time series of heating and cooling consumptions and of internal temperatures associated with each set of parameters are obtained from the physical model. The objective is to build a simplified energy model and to calibrate this model using the input and output data. This will allow then to use the metamodel to propose new tuned parameters to reduce energy consumptions with a given comfort. The metric considered for the challenge is the MSE (mean squared error).
Rakuten France Multimodal Product Data Classification
The goal of this data challenge is large-scale multimodal (text and image) product data classification into _product type codes_. For example, in Rakuten France catalog, a product with a French designation or title _Klarstein Présentoir 2 Montres Optique Fibre_ associated with an image and sometimes with an additional description. This product is categorized under the _1500_ product type code. There are other products with different titles, images and with possible descriptions, which are under the same product type code. Given these information on the products, like the example above, this challenge proposes to model a classifier to classify the products into its corresponding product type code.
NLP applied to judicial decisions parsing
At Predilex, we have “jurisprudence” data as text files and we want to build an algorithm to automate the extraction of the relevant information. In this challenge, we want to extract from "jurisprudence" the sex of the victim, the date of the accident and the date of the consolidation of the injuries.
Wind Power forecasting for the day-ahead energy market
The goal of this challenge is to predict the energy production of six WF owned by CNR. Each WF production will be individually predicted, using meteorological forecasts as input. Predictions will focus on the day-ahead energy production (hourly production forecasts from day D+1 00h to day D+2 00h).
Stock Return Prediction

The proposed challenge aims at predicting the return of a stock in the US market using historical data over a recent period of 20 days. The one-day return of a stock jj on day tt with price PjtP_j^t (adjusted from dividends and stock splits) is given by:

Rjt=PjtPjt11 R_j^t = \frac{P_j^t}{P_j^{t-1}} - 1

In this challenge, we consider the residual stock return, which corresponds to the return of a stock without the market impact. Historical data are composed of residual stock returns and relative volumes, sampled each day during the 20 last business days (approximately one month). The relative volume Vjt\mathcal V_j^t at time tt of a stock jj among the nn stocks is defined by:

Vˉjt=Vtmedian({Vjt1,,Vjt20})Vjt=Vˉjt1ni=1nVˉit \begin{aligned} \bar{V}_j^t &= \frac{V^t}{\mathrm{median(\{ V_j^{t-1}, \dots, V_j^{t-20}\})}} \\ \mathcal V_j^t &= \bar{V}_j^t - \frac{1}{n} \sum_{i=1}^n \bar{V}_i^t \end{aligned}

where VjtV_j^t is the volume at time tt of a stock jj . We also give additional information about each stock such as its industry and sector.

The metric considered is the accuracy of the predicted residual stock return sign.

Smart meter is coming
The goal is to train an algorithm to replace many monitoring systems which are too intrusive and too expensive. This challenge is known as NILM (Nonintrusive load monitoring) or NIALM (Nonintrusive appliance load monitoring). The aim of the challenge is to find the part of electric consumption in one household dedicated to 4 appliances (washing machine, fridge freezer, TV, kettle). There are no time constraints. The past and the future are known.
Defect Prediction on production lines

The goal of the challenge is to predict defect on starter motor production lines. During production samples assembly, different values (torques, angles ...) are measured on different mounting stations. At the end of the line, additional measures are performed on two test benches in order to isolate defects. As a result, samples are tagged ‘OK’, ‘KO’. We would like to design a model that could identify such defects before the test bench step.