High-confidence predictions under adversarial uncertainty

We study the setting in which the bits of an unknown infinite binary sequence x are revealed sequentially to an observer. We show that very limited assumptions about x allow one to make successful predictions about unseen bits of x. First, we study the problem of successfully predicting a single 0 from among the bits of x. In our model we have only one chance to make a prediction, but may do so at a time of our choosing. This model is applicable to a variety of situations in which we want to perform an action of fixed duration, and need to predict a "safe" time-interval to perform it.

Letting N_t denote the number of 1s among the first t bits of x, we say that x is "ε-weakly sparse" if liminf (N_t/t) ≤ ε. Our main result is a randomized algorithm that, given any ε-weakly sparse sequence x, predicts a 0 of x with success probability as close as desired to 1 -- ε. Thus we can perform this task with essentially the same success probability as under the much stronger assumption that each bit of x takes the value 1 independently with probability ε.

We apply this result to show how to successfully predict a bit (0 or 1) under a broad class of possible assumptions on the sequence x. The assumptions are stated in terms of the behavior of a finite automaton M reading the bits of x. We also propose and solve a variant of the well-studied "ignorant forecasting" problem. For every ε > 0, we give a randomized forecasting algorithm S_ε that, given sequential access to a binary sequence x, makes a prediction of the form: "A p fraction of the next N bits will be 1s." (The algorithm gets to choose p, N, and the time of the prediction.) For any fixed sequence x, the forecast fraction p is accurate to within ±ε with probability 1 − ε.