How to derive first partial derivative for expected return of group of wagers

58 Views Asked by At

As stated in the title, I'm attempting to solve for the first partial derivative for the expected return of a group of wagers in a parimutuel pool.

Wikipedia describes parimutuel betting as "a betting system in which all bets of a particular type are placed together in a pool; taxes and the "house-take" or "vigorish" are deducted, and payoff odds are calculated by sharing the pool among all winning bets."1.

The expected return of a series of potential bets into a parimutuel pool with $N$ combinations can then be calculated by the formula:

$$\sum_{i=1}^N(\frac{Q(W + \sum_{j=1}^Nb_j)}{w_i + b_i} - 1)b_ip_i - (b_i(1 - p_i))$$

where $N$ is the number of combinations, $Q$ is the money returned to bettors by the house after taxes and vig (expressed as a percentage), $W$ is the total money wagered into the pool prior to our proposed bets, $w_i$ and $b_i$ are the money wagered on combination $i$ by the public (not including our wager) and on our wager, respectively (both values are in the range 0+), and $p_i$ is the probability that combination $i$ is the winner.

My first attempt at the partial derivative with respect to $b_i$, after pushing the $b_ip_i$ term into the quotient, was:

$$\frac{Q(Wp_i+\sum_{j\ne i}^N(b_j)p_i+2b_ip_i)(w_i+b_i)-Q(Wb_ip_i+\sum_{j=1}^N(b_j)b_ip_i))}{(w_i+b_i)^2}-1$$

However, I checked the gradient using python's scipy package (scipy.optimize.check_grad), which compares the derived gradient with the numerically estimated gradient. The results were significantly different. After some attempts, I realized I calculated the gradient of the quotient

$$Q(Wb_ip_i + (\sum_{j=1}^Nb_j)b_ip_i)$$

incorrectly. With some help from friends, I was able to solve it as

$$Q(Wp_i+\sum_{j=1}^N(b_jp_i+b_jp_j))$$

However, when I replaced that in the application of the quotient rule from my original attempt, the check_grad result was still significantly different.

At this point, I'm stuck and wondering if pushing the $b_ip_i$ term into the quotient was appropriate given the interaction of $b_i$ in the summation. I'm also wondering if applying the quotient rule to the denominator without considering both $i=j$ and $i \ne j$, is the issue. Any pointers or help would be greatly appreciated!

1

There are 1 best solutions below

0
On BEST ANSWER

After bashing my head against a wall for days, I think I've figured it out. The solution I've coded up passes the scipy check_grad test, so I'm fairly confident it's accurate. However, I'm not confident in my ability to convert that to a proper latex formula, so apologies if what's posted below is inaccurate. Please let me know and I'll modify.

It helped to write out some explicit cases for a small number of bets i.e. $N=2$. From there it was clear that what I was missing were a given bet's impact on the ER when considering a separate bet. The $i=k$ vs $i\ne k$ case. There were a couple S/O posts that were useful there (1)(2). The solution is:

$$\sum_{i=1}^N[Q(\frac{(Wp_i+\sum_{j=1}^N(b_jp_i)+b_ip_i)(w_i+b_i)-(Wb_ip_i+\sum_{j=1}^N(b_j)(b_ip_i))}{(w_i + b_i)^2}+\sum_{j\ne i}^N(\frac{b_jp_j}{w_j + b_j})) - 1]$$