Lag correlation between two time series is the correlation shifted in time relative to one another. Existing work focuses on two computation models, landmark (where the lag correlation is computed over the entire stream) and sliding window (where the lag correlation is computed over the current window). However, these models may suffer from problems like result freshness (e.g., perished items in the landmark model) and parametric tuning (e.g., setting a proper length in the sliding window model). In this work, we attempt to analyze the lag correlation which is computed based on flexible sliding windows. In view of that, a new query called RLC (ranking lag correlations with flexible sliding windows in data streams) is proposed. The key challenge in answering the RLC query is that the number of windows to be analyzed will grow quadratically with the length of the stream, resulting a quadratic computation cost. To boost the computation, we employ the running sum and the geometric probing techniques to facilitate the query processing. We also present an approximate solution that further reduces the computation cost with an acceptable error rate in practice. The extensive experiments verify the efficiency of our proposed methods. We also demonstrate some lag correlations discovered from real datasets to show the practicality of this work.