23 | 25 |
new file mode 100644 |
... | ... |
@@ -0,0 +1,165 @@ |
1 |
+"""Tools for dealing with bookmaker's odds and implied probabilities""" |
|
2 |
+ |
|
3 |
+from math import sqrt |
|
4 |
+from typing import Sequence |
|
5 |
+ |
|
6 |
+import scipy.optimize |
|
7 |
+import toolz |
|
8 |
+ |
|
9 |
+ |
|
10 |
+# -- Adjusting implied probabilities to true probabilities |
|
11 |
+ |
|
12 |
+_adjust_methods = {} |
|
13 |
+ |
|
14 |
+ |
|
15 |
+@toolz.curry |
|
16 |
+def add_method(name: str, f): |
|
17 |
+ _adjust_methods[name] = f |
|
18 |
+ return f |
|
19 |
+ |
|
20 |
+ |
|
21 |
+def overround(odds: Sequence[float]) -> float: |
|
22 |
+ return sum(odds) - 1 |
|
23 |
+ |
|
24 |
+ |
|
25 |
+def validate_odds(odds: Sequence[float]) -> None: |
|
26 |
+ if len(odds) < 2: |
|
27 |
+ raise ValueError(f"Expected at least 2 outcomes, but got {len(odds)}") |
|
28 |
+ if any(x >= 1 for x in odds): |
|
29 |
+ raise ValueError(f"Expected all outcomes to have an implied probability < 1") |
|
30 |
+ r = overround(odds) |
|
31 |
+ if r < 0: |
|
32 |
+ raise ValueError(f"Odds {odds} have a negative overround ({r})") |
|
33 |
+ elif r > 0.5: |
|
34 |
+ raise ValueError(f"Odds {odds} have excessive overround ({r})") |
|
35 |
+ |
|
36 |
+ |
|
37 |
+def adjust_odds(odds: Sequence[float], method="power") -> list[float]: |
|
38 |
+ """Return true odds given bookmaker odds. |
|
39 |
+ |
|
40 |
+ Parameters |
|
41 |
+ ---------- |
|
42 |
+ odds |
|
43 |
+ The odds of each outcome, given as implied probabilites. |
|
44 |
+ |
|
45 |
+ The implied probabilities associated with the odds quoted by a bookmaker |
|
46 |
+ sum to a number > 1. This excess is the "overround" of the odds. There |
|
47 |
+ are a number of models |
|
48 |
+ |
|
49 |
+ Notes |
|
50 |
+ ----- |
|
51 |
+ All implemented methods are based on the descriptions in the paper by |
|
52 |
+ S. Clarke in the references section. |
|
53 |
+ |
|
54 |
+ References |
|
55 |
+ ---------- |
|
56 |
+ S. Clarke, S. Kovalchik, M. Ingram |
|
57 |
+ Adjusting Bookmaker's odds to Allow for Overround. |
|
58 |
+ """ |
|
59 |
+ try: |
|
60 |
+ f = _adjust_methods[method] |
|
61 |
+ except KeyError: |
|
62 |
+ raise ValueError( |
|
63 |
+ f"Unknown method '{method}', expected one of {list(_adjust_methods)}" |
|
64 |
+ ) |
|
65 |
+ validate_odds(odds) |
|
66 |
+ return f(odds) |
|
67 |
+ |
|
68 |
+ |
|
69 |
+@add_method("additive") |
|
70 |
+def adjust_odds_additive(odds: Sequence[float]) -> list[float]: |
|
71 |
+ """Adjust bookmaker odds using the additive method. |
|
72 |
+ |
|
73 |
+ In this method the overround is assumed to be split equality between |
|
74 |
+ the different outcomes, regardless of their relative implied probabilities. |
|
75 |
+ """ |
|
76 |
+ y = overround(odds) / len(odds) |
|
77 |
+ return [x - y for x in odds] |
|
78 |
+ |
|
79 |
+ |
|
80 |
+@add_method("multiplicative") |
|
81 |
+def adjust_odds_multiplicative(odds: Sequence[float]) -> list[float]: |
|
82 |
+ """Adjust bookmaker odds using the multiplicative method. |
|
83 |
+ |
|
84 |
+ In this method the overround is assumed to be split among the |
|
85 |
+ different outcomes in proportion to their implied probabilities. |
|
86 |
+ """ |
|
87 |
+ s = sum(odds) |
|
88 |
+ r = overround(odds) |
|
89 |
+ return [x - (r * x / s) for x in odds] |
|
90 |
+ |
|
91 |
+ |
|
92 |
+@add_method("power") |
|
93 |
+def adjust_odds_power(odds: Sequence[float]) -> list[float]: |
|
94 |
+ """Adjust bookmaker odds using the power method. |
|
95 |
+ |
|
96 |
+ In this method the true probabilities are assumed to be |
|
97 |
+ related to the implied probabilities by a power law, |
|
98 |
+ i.e. p_i = π_i^k, for some fixed k. The k is then |
|
99 |
+ chosen such that the true probabilities sum to 1. |
|
100 |
+ |
|
101 |
+ Notes |
|
102 |
+ ----- |
|
103 |
+ The 'k' parameter for the power method is computed by |
|
104 |
+ solving the nonlinear equation Σ_i π_i^k = 1 using |
|
105 |
+ a root-finding algorithm. |
|
106 |
+ |
|
107 |
+ References |
|
108 |
+ ---------- |
|
109 |
+ S.R. Clarke |
|
110 |
+ Adjusting true odds to allow for vigorish |
|
111 |
+ """ |
|
112 |
+ # This bracketing interval should be sufficient for reasonable overrounds |
|
113 |
+ # (not > 50%) and reasonable distribution of implied probability between |
|
114 |
+ # the different outcomes. |
|
115 |
+ try: |
|
116 |
+ k = scipy.optimize.root_scalar( |
|
117 |
+ lambda k: sum(x ** k for x in odds) - 1, |
|
118 |
+ bracket=[0.9, 3], |
|
119 |
+ xtol=1e-14, |
|
120 |
+ ).root |
|
121 |
+ except ValueError as e: |
|
122 |
+ if "different signs" in str(e): |
|
123 |
+ raise ValueError(f"Bracketing interval too small for odds {odds}") from None |
|
124 |
+ |
|
125 |
+ return [x ** k for x in odds] |
|
126 |
+ |
|
127 |
+ |
|
128 |
+@add_method("shin") |
|
129 |
+def adjust_odds_shin(odds): |
|
130 |
+ """Adjust bookmaker odds using the Shin method. |
|
131 |
+ |
|
132 |
+ In this method we assume a fraction z of sharp bettors |
|
133 |
+ and adjust the true probabilities accordingly. |
|
134 |
+ |
|
135 |
+ Notes |
|
136 |
+ ----- |
|
137 |
+ The 'z' parameter for the Shin method is computed |
|
138 |
+ by solving a nonlinear equation using a root-finding algorithm. |
|
139 |
+ |
|
140 |
+ References |
|
141 |
+ ---------- |
|
142 |
+ H.S. Shin |
|
143 |
+ Prices of State Contingent Claims with Insider Traders |
|
144 |
+ and the Favorite-Longshot Bias |
|
145 |
+ """ |
|
146 |
+ s = sum(odds) |
|
147 |
+ n = len(odds) |
|
148 |
+ |
|
149 |
+ def residual(z): |
|
150 |
+ return ( |
|
151 |
+ sum(sqrt(z ** 2 + 4 * (1 - z) * pi ** 2 / s) for pi in odds) |
|
152 |
+ - 2 |
|
153 |
+ - (n - 2) * z |
|
154 |
+ ) |
|
155 |
+ |
|
156 |
+ # 'z' is in the interval [0, 1). We set the upper limit for the |
|
157 |
+ # bracketing interval to just below 1 to exclude a (spurious) solution at z=1. |
|
158 |
+ try: |
|
159 |
+ z = scipy.optimize.root_scalar(residual, bracket=[0, 0.999], xtol=1e-14).root |
|
160 |
+ except ValueError: |
|
161 |
+ raise ValueError(f"Shin method failed to converge for odds {odds}") from None |
|
162 |
+ |
|
163 |
+ return [ |
|
164 |
+ (sqrt(z ** 2 + 4 * (1 - z) * pi ** 2 / s) - z) / (2 * (1 - z)) for pi in odds |
|
165 |
+ ] |