T9Sim RTB Data Simulator: Generating, Censoring and Benchmarking DSP/MMP/SSP Adtech Data Layers

Simulated real-time-bidding (RTB) auction data for mobile-game advertising. Unlike a real advertising log, every auction here records user action, conversion, wininng prices and other latent features, in the auctions the advertiser won and lost. What is in the deposit 10 datasets of 10 million auctions each, seeds 90213 to 90222, plus a 1-million-auction sample for initial inspection. Files are Apache Parquet, 55 columns per auction, covering 1) observable features such as the bid request context, 2) price and auction features, such as winning price and bid rival density, 3) outcome features, including clicks, installs and 90-day spending, and 4) latent features from the pools the auctions are constructed from, user, app, campaign and rival pools. 23 columns hold ground truth that no real dataset could contain, including each auction's true expected value. These columns are used for evaluation, never for training. About T9Sim This deposit is a data and code snapshot accompanying T9Sim, a data generator and benchmark environment for advertising auctions. T9Sim generates one ground-truth stream of auctions, censors it into four views that mimic what different players in the adtech chain can see, then benchmarks bidding algorithms trained by these four views of data, namely C1-C4. Each view approximates what an adtech business type may see. C1 is a demand side platform (DSP) alone. C2 is a DSP with a mobile measurement partner (MMP), which brings conversion outcomes on every auction. C3 is a DSP with a supply side platform (SSP), which brings winning prices and rival counts. C4 is a DSP with both. Because the auctions stay fixed, comparing the views measures what each data layer is worth. How each auction is generated Four pools of entities are drawn once and frozen for a run, namely users, apps, campaigns and rival bidders, each carrying hidden characteristics that no log records. Each auction then selects one user, one app and one campaign, calculates what that impression is truly worth, and determines which rival bidders participate and what they bid. The buying platform submits its bid without observing those rival bids. The auction then resolves, and the user's response is drawn last, on every row, whether or not an advertisement was shown.