Filtering without recursion and some of its uses in financial economics

Battle Count: 50

Highly relevant for quantitative trading. The paper directly addresses robust price estimation from noisy high-frequency trade data, which is fundamental to algorithmic execution, market making, and volatility-based trading strategies. The flat volatility signature at 1-second level enables reliable sub-minute risk measurement. The identification of stale trades (off-exchange reporting) and their directional bias is directly actionable for execution algorithms. The O(T) computational cost makes it feasible for real-time processing of 1.5M+ daily trades. The Huber filter's robustness to infinite-variance noise is critical for heavily traded stocks where linear methods catastrophically fail.

The core algorithm (Algorithm 1) is straightforward: sample lag lengths from the weight distribution, retrieve corresponding data points, compute the sample minimizer. However, practical implementation requires: (1) efficient sampling from truncated geometric/hyperbolic distributions, (2) stratified estimation with exact head computation, (3) iterative reweighted least squares for Huber loss, (4) running order-statistic trees for median computation, (5) FFT convolution for hyperbolic weights, (6) careful handling of end effects and diurnal patterns, (7) multi-stage pipeline (median filter -> noise scale -> Huber filter -> volatility). The paper provides sufficient algorithmic detail but no reference code. Parallelization across t is natural but requires careful memory management for 1.5M+ data points.

The paper provides detailed algorithms (Algorithm 1 for sampling, stratified estimator in Remark 4(d)), specific parameter choices (half-lives, B=100, h=W_t^{-1}(0.999)), and uses publicly available NYSE TAQ data via WRDS. The R package spatstat is referenced for weighted median. Web Appendix covers 54 additional assets. However, no explicit GitHub repository is provided, and full replication requires WRDS access.