def hit_conf(conf,p, lift, interval, consecutives = 1, max_run = float('inf')):
A = Banner(p+(p*lift), BernouliEstimator(alpha = 1, beta = 1))
B = Banner(p, BernouliEstimator(alpha = 1, beta = 1))
time = 0
hit_better = 0
hit_worse = 0
while(hit_better < consecutives and hit_worse < consecutives):
time+=interval
A.run(interval)
B.run(interval)
dist_a = A.estimator.get_distribution()
dist_b = B.estimator.get_distribution()
lift_dist = (dist_a - dist_b)/dist_b
if np.mean(lift_dist > 0) > conf:
hit_better += 1
else:
hit_better = 0
if np.mean(lift_dist < 0) > conf:
hit_worse += 1
else:
hit_worse = 0
if max_run < time:
break
if hit_better < consecutives:
return 0.0, time
else:
return 1.0, time
def estimate_baseline(p, conf, min_effect, alpha = 1, beta = 1):
interval = round((1/p) * 10)
time = interval
A = Banner(p, BernouliEstimator(alpha = 1, beta =1))
A.run(interval)
dist = A.estimator.get_distribution()
m = np.mean(dist)
l = np.percentile(dist, 100*min_effect)
while(m-l > min_effect*l):
time += interval
A.run(interval)
dist = A.estimator.get_distribution()
m = np.mean(dist)
l = np.percentile(dist, 100*min_effect)
return time, m, l
