RTB ripped black box part 1:datacratic ' s RTB algorithms

This article is a discussion of some of the statistical and economic theories used by datacratic . We developed the real time bidding algorithm s. To achieve The advertiser's goal, our algorithm automatically leverages the suboptimal strategy of other advertisers and then looks at the ad's reserve price. We want our partners to understand the technology we use and think it is reasonable. The values of "believing in the black box" are not tenable here.

First , tell the Truth

Assuming that there is now an effect ad business, the total budget is$100,000，CPCIs$1.00。 Like the otherDSP, you need to subscribe to a thousands of bid request per second(ThatAd Call), each request represents an opportunity to show a specific user at a specific ad position. You must bid for a request that matches your targeting criteria within dozens of milliseconds. So how do you decide how much to bid? Most real-time auctions are based onVCGBilling or second-order billing methods. The winner is the highest bidder, but he only pays the second highest bidder's bid.Ad ExchangeThis approach is used because it encourages the bidder to "truly bid (bid truthfully) ". Also, you should direct the value of what you think the show is worth. But how much does this show cost? First target cpc is $1.00 , then a good assumption is that advertisers think at least this click Value $1.00 . We assume that the average clickthrough rate for this promotion plan is 0.05% . So if you win this bid, then there is 0.05% opportunity to get value < Span style= "FONT-FAMILY:CALIBRI;" >$1.00 Click, we will bid and click-through rate, we can think that the value of winning is 0.05 . If you are real bid by auction theory, you should do this ) , you should bid 0.05 . Summarize it as a formula:

Bid = value = TARGETCPC * CTR

This explanation is not too simple, and it does not take advantage of the real-time characteristics of the Adx . We are a predictive analytics company, and we have a few flashy models that can predict a particular ad at a specific time in a specific ad position at a given moment in real time, in a way that predicts a single request. That means you don't need to count the historical CTR of this promotion plan when calculating your bids , we use the pctr model to predict bids:

Bid = value = TARGETCPC * P (click)

The deduction above seems to be fine, but there seems to be nothing special about it. In addition to PCTR this model beyond the content of this article, then what is the secret of bidding? In fact, the formula will only answer if you decide to bid what your bids should be. But when to bid, the formula is not answered. When bidding, in fact you should not just look at the bids, but also look at the cost.

to bid or not to Bid, Or:cost was not Value

As we discussed in Part0 , if you bid on every offer, you will quickly consume $100,000and not stick to 3 months. We'll think more deeply than Part0 : Is it possible for us to bid without randomly selecting the traffic, but choosing some of the best requests for bidding? The best request does not refer to the highest-valued request, and the highest value is only half the best definition. To explain the problem, let's cite an example where a purchase can buy a or B merchandise for two sales. The value of commodity A Two sales is lower than b 10%, but it costs only half of b .

Two items have similar value but different profits

In this case a low value of a is a better commodity, because it is not the value and cost of determining the good or bad, but the difference between the two, which we call profit.

Surplus = Profit = Payoff = Value-cost

The formula above is set up when you win, but what if you don't win? You will not be able to get one click, so the value is 0, but your cost is also 0. So our expectation value is the probability that the above formula Sheng out:

Surplus = (value-cost) * P (Win)

Now suppose that in addition to datacratic 's pctr model, we have a competitive price ( out of the price, which is the price that is fully competitive in the market ) The predictive model tells us the probability of winning each request, so with it, we can calculate the profit.

The surface is a function of bid and value, for any value, the maximum profit is in the bid = Value when

In part 0 We have introduced the goal of getting a budget, we can only bid on some of the traffic, we should only select the portion of the request that is most profitable. means that the variable that controls the bid is no longer the bid probability, and a value that measures how much you want to participate in the bid: the smaller the value, the more selective you will be. You're still adjusting with feedback, but now you're only bidding on the request that you think will make the most profit, so your chances of getting a lower CPC become bigger.

When you take a real bid and Pacewisely, your budget will be spent on the best request, either because the user is likely to click, or because there are currently few DSPs bids, or if no publisher has set a reserve price.

RTB ripped black box part 1:datacratic ' s RTB algorithms