I'm successfully using Welford's method to compute running variance and standard deviation as described many times on Stack Overflow and John D Cook's excellent blog post.
I store the timestamp, count, sum, and running calculation of "Sk" and stdev in a database table PER DAY.
I would like to combine or rollup the daily computed count,sum, and Sk values get a MONTHLY standard deviation.
John D Cook's blog has another post that provides an algorithm to combine two "RunningStats" into 1 (see operator+ method). This works for combining 2 days into 1. I could use this to iterate through the days to combine all days in the month. However, unlike calculating daily stdev where I have a large number of samples that must be dealt with in a streaming fashion, I have access to all the daily data, and would like a single formula to combine all days in the month at once. That would lend itself to creation of a database view.
It does not appear that simply summing up the Sk values and dividing by the total monthly count - 1 produces accurate variance.
Example Data:
DATE, COUNT, SUM, Sk, STDDEV
1-Jun-15, 60, 514, 1556.733336, 5.14
2-Jun-15, 51, 455, 1523.686274, 5.52
3-Jun-15, 61, 556, 1494.196722, 4.99
...
Let x1, ..., xn be the values for one day. Then, as I understand it, Sk is defined (mathematically; the rolling implementation is different) as follows.
Mk = (sum_{i=1}^n xi) / n [Mk = SUM / COUNT]
Sk = sum_{i=1}^n (xi - Mk)^2.
The problem with summing the Sk column is that each value is computed with respect to a different mean, so the overall variance is an underestimate. Instead, we should have a term like
sum_{i=1}^n (xi - (Mk + delta))^2,
which we rewrite in terms of the existing quantities.
sum_{i=1}^n (xi - (Mk + delta))^2
= sum_{i=1}^n (xi - Mk - delta)^2
= sum_{i=1}^n ((xi - Mk)^2 - 2 (xi - Mk) delta + delta^2)
= Sk + n delta^2.
Here delta is the monthly mean minus the daily mean. Once again, I do not warrant numerical stability.
For my work, I need some kind of algorithm with the following input and output:
Input: a set of dates (from the past). Output: a set of weights - one weight per one given date (the sum of all weights = 1).
The basic idea is that the closest date to today's date should receive the highest weight, the second closest date will get the second highest weight, and so on...
Any ideas?
Thanks in advance!
First, for each date in your input set assign the amount of time between the date and today.
For example: the following date set {today, tomorrow, yesterday, a week from today} becomes {0, 1, 1, 7}. Formally: val[i] = abs(today - date[i]).
Second, inverse the values in such a way that their relative weights are reversed. The simplest way of doing so would be: val[i] = 1/val[i].
Other suggestions:
val[i] = 1/val[i]^2
val[i] = 1/sqrt(val[i])
val[i] = 1/log(val[i])
The hardest and most important part is deciding how to inverse the values. Think, what should be the nature of the weights? (do you want noticeable differences between two far away dates, or maybe two far away dates should have pretty equal weights? Do you want a date which is very close to today have an extremely bigger weight or a reasonably bigger weight?).
Note that you should come up with an inverting procedure where you cannot divide by zero. In the example above, dividing by val[i] results in division by zero. One method to avoid division by zero is called smoothing. The most trivial way to "smooth" your data is using the add-one smoothing where you just add one to each value (so today becomes 1, tomorrow becomes 2, next week becomes 8, etc).
Now the easiest part is to normalize the values so that they'll sum up to one.
sum = val[1] + val[2] + ... + val[n]
weight[i] = val[i]/sum for each i
Sort dates and remove dups
Assign values (maybe starting from the farthest date in steps of 10 or whatever you need - these value can be arbitrary, they just reflect order and distance)
Normalize weights to add up to 1
Executable pseudocode (tweakable):
#!/usr/bin/env python
import random, pprint
from operator import itemgetter
# for simplicity's sake dates are integers here ...
pivot_date = 1000
past_dates = set(random.sample(range(1, pivot_date), 5))
weights, stepping = [], 10
for date in sorted(past_dates):
weights.append( (date, stepping) )
stepping += 10
sum_of_steppings = sum([ itemgetter(1)(x) for x in weights ])
normalized = [ (d, (w / float(sum_of_steppings)) ) for d, w in weights ]
pprint.pprint(normalized)
# Example output
# The 'date' closest to 1000 (here: 889) has the highest weight,
# 703 the second highest, and so forth ...
# [(151, 0.06666666666666667),
# (425, 0.13333333333333333),
# (571, 0.2),
# (703, 0.26666666666666666),
# (889, 0.3333333333333333)]
How to weight: just compute the difference of all dates and the current date
x(i) = abs(date(i) - current_date)
you can then use different expression to assign weights:
w(i) = 1/x(i)
w(i) = exp(-x(i))
w(i) = exp(-x(i)^2))
use gaussian distribution - more complicated, do not recommend
Then use normalized weights: w(i)/sum(w(i)) so that the sum is 1.
(Note that the exponential func is always used by statisticians in survival analysis)
The first thing that comes to my mind to to use a geometric series:
http://en.wikipedia.org/wiki/Geometric_series
(1/2)+(1/4)+(1/8)+(1/16)+(1/32)+(1/64)+(1/128)+(1/256)..... sums to one.
Yesterday would be 1/2
2 days ago would be 1/4
and so on
Is is the index for the i-th date.
Assign weights equal to to Ni / D.
D0 is the first date.
Ni is the difference in days between the i-th date and the first date D0.
D is the normalization factor
converts dates to yyyymmddhhmiss format (24 hours), add all these values and the total, divide by the total time, and sort by this value.
declare #data table
(
Date bigint,
Weight float
)
declare #sumTotal decimal(18,2)
insert into #Data (Date)
select top 100
replace(replace(replace(convert(varchar,Datetime,20),'-',''),':',''),' ','')
from Dates
select #sumTotal=sum(Date)
from #Data
update #Data set
Weight=Date/#sumTotal
select * from #Data order by 2 desc
I have a list of 6500 items that I would like to trade or invest in. (Not for real money, but for a certain game.) Each item has 5 numbers that will be used to rank it among the others.
Total quantity of item traded per day: The higher this number, the better.
The Donchian Channel of the item over the last 5 days: The higher this number, the better.
The median spread of the price: The lower this number, the better.
The spread of the 20 day moving average for the item: The lower this number, the better.
The spread of the 5 day moving average for the item: The higher this number, the better.
All 5 numbers have the same 'weight', or in other words, they should all affect the final number in the with the same worth or value.
At the moment, I just multiply all 5 numbers for each item, but it doesn't rank the items the way I would them to be ranked. I just want to combine all 5 numbers into a weighted number that I can use to rank all 6500 items, but I'm unsure of how to do this correctly or mathematically.
Note: The total quantity of the item traded per day and the donchian channel are numbers that are much higher then the spreads, which are more of percentage type numbers. This is probably the reason why multiplying them all together didn't work for me; the quantity traded per day and the donchian channel had a much bigger role in the final number.
The reason people are having trouble answering this question is we have no way of comparing two different "attributes". If there were just two attributes, say quantity traded and median price spread, would (20million,50%) be worse or better than (100,1%)? Only you can decide this.
Converting everything into the same size numbers could help, this is what is known as "normalisation". A good way of doing this is the z-score which Prasad mentions. This is a statistical concept, looking at how the quantity varies. You need to make some assumptions about the statistical distributions of your numbers to use this.
Things like spreads are probably normally distributed - shaped like a normal distribution. For these, as Prasad says, take z(spread) = (spread-mean(spreads))/standardDeviation(spreads).
Things like the quantity traded might be a Power law distribution. For these you might want to take the log() before calculating the mean and sd. That is the z score is z(qty) = (log(qty)-mean(log(quantities)))/sd(log(quantities)).
Then just add up the z-score for each attribute.
To do this for each attribute you will need to have an idea of its distribution. You could guess but the best way is plot a graph and have a look. You might also want to plot graphs on log scales. See wikipedia for a long list.
You can replace each attribute-vector x (of length N = 6500) by the z-score of the vector Z(x), where
Z(x) = (x - mean(x))/sd(x).
This would transform them into the same "scale", and then you can add up the Z-scores (with equal weights) to get a final score, and rank the N=6500 items by this total score. If you can find in your problem some other attribute-vector that would be an indicator of "goodness" (say the 10-day return of the security?), then you could fit a regression model of this predicted attribute against these z-scored variables, to figure out the best non-uniform weights.
Start each item with a score of 0. For each of the 5 numbers, sort the list by that number and add each item's ranking in that sorting to its score. Then, just sort the items by the combined score.
You would usually normalize your data entries to their respective range. Since there is no fixed range for them, you'll have to use a sliding range - or, to keep it simpler, normalize them to the daily ranges.
For each day, get all entries for a given type, get the highest and the lowest of them, determine the difference between them. Let Bottom=value of the lowest, Range=difference between highest and lowest. Then you calculate for each entry (value - Bottom)/Range, which will result in something between 0.0 and 1.0. These are the numbers you can continue to work with, then.
Pseudocode (brackets replaced by indentation to make easier to read):
double maxvalues[5];
double minvalues[5];
// init arrays with any item
for(i=0; i<5; i++)
maxvalues[i] = items[0][i];
minvalues[i] = items[0][i];
// find minimum and maximum values
foreach (items as item)
for(i=0; i<5; i++)
if (minvalues[i] > item[i])
minvalues[i] = item[i];
if (maxvalues[i] < item[i])
maxvalues[i] = item[i];
// now scale them - in this case, to the range of 0 to 1.
double scaledItems[sizeof(items)][5];
double t;
foreach(i=0; i<5; i++)
double delta = maxvalues[i] - minvalues[i];
foreach(j=sizeof(items)-1; j>=0; --j)
scaledItems[j][i] = (items[j][i] - minvalues[i]) / delta;
// linear normalization
something like that. I'll be more elegant with a good library (STL, boost, whatever you have on the implementation platform), and the normalization should be in a separate function, so you can replace it with other variations like log() as the need arises.
Total quantity of item traded per day: The higher this number, the better. (a)
The Donchian Channel of the item over the last 5 days: The higher this number, the better. (b)
The median spread of the price: The lower this number, the better. (c)
The spread of the 20 day moving average for the item: The lower this number, the better. (d)
The spread of the 5 day moving average for the item: The higher this number, the better. (e)
a + b -c -d + e = "score" (higher score = better score)
I'm trying to sort a bunch of products by customer ratings using a 5 star system. The site I'm setting this up for does not have a lot of ratings and continue to add new products so it will usually have a few products with a low number of ratings.
I tried using average star rating but that algorithm fails when there is a small number of ratings.
Example a product that has 3x 5 star ratings would show up better than a product that has 100x 5 star ratings and 2x 2 star ratings.
Shouldn't the second product show up higher because it is statistically more trustworthy because of the larger number of ratings?
Prior to 2015, the Internet Movie Database (IMDb) publicly listed the formula used to rank their Top 250 movies list. To quote:
The formula for calculating the Top Rated 250 Titles gives a true Bayesian estimate:
weighted rating (WR) = (v ÷ (v+m)) × R + (m ÷ (v+m)) × C
where:
R = average for the movie (mean)
v = number of votes for the movie
m = minimum votes required to be listed in the Top 250 (currently 25000)
C = the mean vote across the whole report (currently 7.0)
For the Top 250, only votes from regular voters are considered.
It's not so hard to understand. The formula is:
rating = (v / (v + m)) * R +
(m / (v + m)) * C;
Which can be mathematically simplified to:
rating = (R * v + C * m) / (v + m);
The variables are:
R – The item's own rating. R is the average of the item's votes. (For example, if an item has no votes, its R is 0. If someone gives it 5 stars, R becomes 5. If someone else gives it 1 star, R becomes 3, the average of [1, 5]. And so on.)
C – The average item's rating. Find the R of every single item in the database, including the current one, and take the average of them; that is C. (Suppose there are 4 items in the database, and their ratings are [2, 3, 5, 5]. C is 3.75, the average of those numbers.)
v – The number of votes for an item. (To given another example, if 5 people have cast votes on an item, v is 5.)
m – The tuneable parameter. The amount of "smoothing" applied to the rating is based on the number of votes (v) in relation to m. Adjust m until the results satisfy you. And don't misinterpret IMDb's description of m as "minimum votes required to be listed" – this system is perfectly capable of ranking items with less votes than m.
All the formula does is: add m imaginary votes, each with a value of C, before calculating the average. In the beginning, when there isn't enough data (i.e. the number of votes is dramatically less than m), this causes the blanks to be filled in with average data. However, as votes accumulates, eventually the imaginary votes will be drowned out by real ones.
In this system, votes don't cause the rating to fluctuate wildly. Instead, they merely perturb it a bit in some direction.
When there are zero votes, only imaginary votes exist, and all of them are C. Thus, each item begins with a rating of C.
See also:
A demo. Click "Solve".
Another explanation of IMDb's system.
An explanation of a similar Bayesian star-rating system.
Evan Miller shows a Bayesian approach to ranking 5-star ratings:
where
nk is the number of k-star ratings,
sk is the "worth" (in points) of k stars,
N is the total number of votes
K is the maximum number of stars (e.g. K=5, in a 5-star rating system)
z_alpha/2 is the 1 - alpha/2 quantile of a normal distribution. If you want 95% confidence (based on the Bayesian posterior distribution) that the actual sort criterion is at least as big as the computed sort criterion, choose z_alpha/2 = 1.65.
In Python, the sorting criterion can be calculated with
def starsort(ns):
"""
http://www.evanmiller.org/ranking-items-with-star-ratings.html
"""
N = sum(ns)
K = len(ns)
s = list(range(K,0,-1))
s2 = [sk**2 for sk in s]
z = 1.65
def f(s, ns):
N = sum(ns)
K = len(ns)
return sum(sk*(nk+1) for sk, nk in zip(s,ns)) / (N+K)
fsns = f(s, ns)
return fsns - z*math.sqrt((f(s2, ns)- fsns**2)/(N+K+1))
For example, if an item has 60 five-stars, 80 four-stars, 75 three-stars, 20 two-stars and 25 one-stars, then its overall star rating would be about 3.4:
x = (60, 80, 75, 20, 25)
starsort(x)
# 3.3686975120774694
and you can sort a list of 5-star ratings with
sorted([(60, 80, 75, 20, 25), (10,0,0,0,0), (5,0,0,0,0)], key=starsort, reverse=True)
# [(10, 0, 0, 0, 0), (60, 80, 75, 20, 25), (5, 0, 0, 0, 0)]
This shows the effect that more ratings can have upon the overall star value.
You'll find that this formula tends to give an overall rating which is a bit
lower than the overall rating reported by sites such as Amazon, Ebay or Wal-mart
particularly when there are few votes (say, less than 300). This reflects the
higher uncertainy that comes with fewer votes. As the number of votes increases
(into the thousands) all overall these rating formulas should tend to the
(weighted) average rating.
Since the formula only depends on the frequency distribution of 5-star ratings
for the item itself, it is easy to combine reviews from multiple sources (or,
update the overall rating in light of new votes) by simply adding the frequency
distributions together.
Unlike the IMDb formula, this formula does not depend on the average score
across all items, nor an artificial minimum number of votes cutoff value.
Moreover, this formula makes use of the full frequency distribution -- not just
the average number of stars and the number of votes. And it makes sense that it
should since an item with ten 5-stars and ten 1-stars should be treated as
having more uncertainty than (and therefore not rated as highly as) an item with
twenty 3-star ratings:
In [78]: starsort((10,0,0,0,10))
Out[78]: 2.386028063783418
In [79]: starsort((0,0,20,0,0))
Out[79]: 2.795342687927806
The IMDb formula does not take this into account.
See this page for a good analysis of star-based rating systems, and this one for a good analysis of upvote-/downvote- based systems.
For up and down voting you want to estimate the probability that, given the ratings you have, the "real" score (if you had infinite ratings) is greater than some quantity (like, say, the similar number for some other item you're sorting against).
See the second article for the answer, but the conclusion is you want to use the Wilson confidence. The article gives the equation and sample Ruby code (easily translated to another language).
Well, depending on how complex you want to make it, you could have ratings additionally be weighted based on how many ratings the person has made, and what those ratings are. If the person has only made one rating, it could be a shill rating, and might count for less. Or if the person has rated many things in category a, but few in category b, and has an average rating of 1.3 out of 5 stars, it sounds like category a may be artificially weighed down by the low average score of this user, and should be adjusted.
But enough of making it complex. Let’s make it simple.
Assuming we’re working with just two values, ReviewCount and AverageRating, for a particular item, it would make sense to me to look ReviewCount as essentially being the “reliability” value. But we don’t just want to bring scores down for low ReviewCount items: a single one-star rating is probably as unreliable as a single 5 star rating. So what we want to do is probably average towards the middle: 3.
So, basically, I’m thinking of an equation something like X * AverageRating + Y * 3 = the-rating-we-want. In order to make this value come out right we need X+Y to equal 1. Also we need X to increase in value as ReviewCount increases...with a review count of 0, x should be 0 (giving us an equation of “3”), and with an infinite review count X should be 1 (which makes the equation = AverageRating).
So what are X and Y equations? For the X equation want the dependent variable to asymptotically approach 1 as the independent variable approaches infinity. A good set of equations is something like:
Y = 1/(factor^RatingCount)
and (utilizing the fact that X must be equal to 1-Y)
X = 1 – (1/(factor^RatingCount)
Then we can adjust "factor" to fit the range that we're looking for.
I used this simple C# program to try a few factors:
// We can adjust this factor to adjust our curve.
double factor = 1.5;
// Here's some sample data
double RatingAverage1 = 5;
double RatingCount1 = 1;
double RatingAverage2 = 4.5;
double RatingCount2 = 5;
double RatingAverage3 = 3.5;
double RatingCount3 = 50000; // 50000 is not infinite, but it's probably plenty to closely simulate it.
// Do the calculations
double modfactor = Math.Pow(factor, RatingCount1);
double modRating1 = (3 / modfactor)
+ (RatingAverage1 * (1 - 1 / modfactor));
double modfactor2 = Math.Pow(factor, RatingCount2);
double modRating2 = (3 / modfactor2)
+ (RatingAverage2 * (1 - 1 / modfactor2));
double modfactor3 = Math.Pow(factor, RatingCount3);
double modRating3 = (3 / modfactor3)
+ (RatingAverage3 * (1 - 1 / modfactor3));
Console.WriteLine(String.Format("RatingAverage: {0}, RatingCount: {1}, Adjusted Rating: {2:0.00}",
RatingAverage1, RatingCount1, modRating1));
Console.WriteLine(String.Format("RatingAverage: {0}, RatingCount: {1}, Adjusted Rating: {2:0.00}",
RatingAverage2, RatingCount2, modRating2));
Console.WriteLine(String.Format("RatingAverage: {0}, RatingCount: {1}, Adjusted Rating: {2:0.00}",
RatingAverage3, RatingCount3, modRating3));
// Hold up for the user to read the data.
Console.ReadLine();
So you don’t bother copying it in, it gives this output:
RatingAverage: 5, RatingCount: 1, Adjusted Rating: 3.67
RatingAverage: 4.5, RatingCount: 5, Adjusted Rating: 4.30
RatingAverage: 3.5, RatingCount: 50000, Adjusted Rating: 3.50
Something like that? You could obviously adjust the "factor" value as needed to get the kind of weighting you want.
You could sort by median instead of arithmetic mean. In this case both examples have a median of 5, so both would have the same weight in a sorting algorithm.
You could use a mode to the same effect, but median is probably a better idea.
If you want to assign additional weight to the product with 100 5-star ratings, you'll probably want to go with some kind of weighted mode, assigning more weight to ratings with the same median, but with more overall votes.
If you just need a fast and cheap solution that will mostly work without using a lot of computation here's one option (assuming a 1-5 rating scale)
SELECT Products.id, Products.title, avg(Ratings.score), etc
FROM
Products INNER JOIN Ratings ON Products.id=Ratings.product_id
GROUP BY
Products.id, Products.title
ORDER BY (SUM(Ratings.score)+25.0)/(COUNT(Ratings.id)+20.0) DESC, COUNT(Ratings.id) DESC
By adding in 25 and dividing by the total ratings + 20 you're basically adding 10 worst scores and 10 best scores to the total ratings and then sorting accordingly.
This does have known issues. For example, it unfairly rewards low-scoring products with few ratings (as this graph demonstrates, products with an average score of 1 and just one rating score a 1.2 while products with an average score of 1 and 1k+ ratings score closer to 1.05). You could also argue it unfairly punishes high-quality products with few ratings.
This chart shows what happens for all 5 ratings over 1-1000 ratings:
http://www.wolframalpha.com/input/?i=Plot3D%5B%2825%2Bxy%29/%2820%2Bx%29%2C%7Bx%2C1%2C1000%7D%2C%7By%2C0%2C6%7D%5D
You can see the dip upwards at the very bottom ratings, but overall it's a fair ranking, I think. You can also look at it this way:
http://www.wolframalpha.com/input/?i=Plot3D%5B6-%28%2825%2Bxy%29/%2820%2Bx%29%29%2C%7Bx%2C1%2C1000%7D%2C%7By%2C0%2C6%7D%5D
If you drop a marble on most places in this graph, it will automatically roll towards products with both higher scores and higher ratings.
Obviously, the low number of ratings puts this problem at a statistical handicap. Never the less...
A key element to improving the quality of an aggregate rating is to "rate the rater", i.e. to keep tabs of the ratings each particular "rater" has supplied (relative to others). This allows weighing their votes during the aggregation process.
Another solution, more of a cope out, is to supply the end-users with a count (or a range indication thereof) of votes for the underlying item.
One option is something like Microsoft's TrueSkill system, where the score is given by mean - 3*stddev, where the constants can be tweaked.
After look for a while, I choose the Bayesian system.
If someone is using Ruby, here a gem for it:
https://github.com/wbotelhos/rating
I'd highly recommend the book Programming Collective Intelligence by Toby Segaran (OReilly) ISBN 978-0-596-52932-1 which discusses how to extract meaningful data from crowd behaviour. The examples are in Python, but its easy enough to convert.