diff --git a/util/rank_bootstrap.py b/util/rank_bootstrap.py
new file mode 100755
index 0000000..e93a25a
--- /dev/null
+++ b/util/rank_bootstrap.py
@@ -0,0 +1,57 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+from __future__ import unicode_literals
+import sys
+from argparse import ArgumentParser
+
+import pandas as pd
+import numpy.random as rnd
+
+from tgen.debug import exc_info_hook
+sys.excepthook = exc_info_hook
+
+
+def pairwise_bootstrap(data_a, data_b, label_a, label_b, iters):
+
+    assert(len(data_a) == len(data_b))
+
+    a_better, b_better, ties = 0, 0, 0
+    for i in xrange(iters):
+        sample = rnd.randint(0, len(data_a), len(data_a))
+        sum_a_good = sum(data_a[i] for i in sample)
+        sum_b_good = sum(data_b[i] for i in sample)
+
+        if sum_a_good > sum_b_good:
+            a_better += 1
+        elif sum_b_good > sum_a_good:
+            b_better += 1
+        else:
+            ties += 1
+
+    print (('%s better: %d (%2.2f) | %s better: %d (%2.2f) | ties: %d (%2.2f)') %
+           (label_a, a_better, float(a_better) / iters * 100,
+            label_b, b_better, float(b_better) / iters * 100,
+            ties, float(ties) / iters * 100,))
+
+
+def main(args):
+    rnd.seed(1206)
+    data_a = pd.read_csv(args.results_file_a, encoding='UTF-8', sep=b'\t', index_col=None)
+    data_b = pd.read_csv(args.results_file_b, encoding='UTF-8', sep=b'\t', index_col=None)
+
+    pairwise_bootstrap(list(data_a[args.target_column + '_rank_ok']),
+                       list(data_b[args.target_column + '_rank_ok']),
+                       args.results_file_a, args.results_file_b,
+                       args.bootstrap_iters)
+
+
+if __name__ == '__main__':
+    ap = ArgumentParser()
+    ap.add_argument('-b', '--bootstrap-iters', type=int, default=1000)
+    ap.add_argument('-t', '--target-column', type=str, default='quality')
+    ap.add_argument('results_file_a', type=str)
+    ap.add_argument('results_file_b', type=str)
+    args = ap.parse_args()
+
+    main(args)