S.A.D. Predictions - Eights 2011

Iteration#8: 4th June

These are the updated scores and predictions (using the original Wednesday starting positions) using the SAD BaStaRD program to incorporate results after three days of racing.

I imagine the main interest will be on seeing how your crew estimates have changed after yesterday's results, in which case click on the crew name and the left plot shows the evolution of your SAD score

Original contributions from NP, JJ, SP, MV, LL, PC, JC, PM, BW, IG, ST, CP, AD, KS, LG, FM, TC, JS. The updated Contributor Statistics now start show the skill of each contributor in terms of the agreement between their estimates and those derived from SAD BaStaRD, although still weighted towards the original mean scores. However, these average errors will be used to weight the contibutors estimates in the next iteration before rerunning SAD BaStaRD.


Predictions for the Saturday finishing order using the Statistical Analysis of Divisions (SAD) program based on collective estimates of boat speeds (top divisions only). The following tables show the Wednesday starting positions of each crew, its score and average predicted change in position by the end of the week, followed by the crews most likely to occupy that position at the end of the week, listed in order of decreasing probability.

Click on crew name for graphics showing evolution of SAD score with each iteration and spread of Predicted Net Changes over the week.

The Evolution of SAD Score shows score each iteration as a red cross, the vertical extent representing the uncertainty. The previous Torpids score used as an initial estimate and its associate uncertainty is shown as a green box. The Norm for the crew's starting position is shown as a blue dashed line. In the absence of any new information, each iteration the score will relax towards the "Norm" and the error bar expand towards ±2. If new estimates have been supplied the error bar will get smaller.

The Predicted Net Change histogram plot shows the probability of different net gains in position over the week (blue columns for rises, red columns for falls, white for remaining the same). The dashed line shows the mean change in position averaged over all possible outcomes (all the information for these histograms is extracted from the tables below).


[Prev]

Men's Divisions 1-3

Pos Start Crew Score Av.Rise Finishing Crew
Division 1
1Christ Church 5.6 -2.5 Orie 94% Pemb 5% Chri 1%
2Oriel 6.5 +0.9 Pemb 86% Chri 8% Orie 5% Univ 0% Hert 0% SCat 0%
3Pembroke 6.3 +0.9 Chri 52% Univ 27% Pemb 9% Hert 8% SCat 3% Orie 1% Ball 1%
4University 5.8 -0.5 Wolf 59% Univ 16% Chri 13% Hert 7% SCat 3% Ball 1% Pemb 0% Orie 0%
5Hertford 5.8 -0.7 Univ 38% Chri 25% Wolf 18% Hert 13% SCat 4% Ball 2% Pemb 0% Orie 0%
6St Catherine's 5.8 -0.8 Hert 52% SCat 15% Univ 13% Wolf 13% Ball 5% Magd 1% SEdm 0% Pemb 0%
7Balliol 5.8 -0.9 SCat 55% Ball 15% Hert 14% Wolf 7% Magd 4% Univ 4% SEdm 1% Worc 0%
8Wolfson 6.5 +3.2 Ball 59% SCat 16% Magd 13% Hert 5% Wolf 3% SEdm 2% Univ 1% Worc 0%
9Magdalen 5.8 -0.5 SJoh 70% Magd 18% Ball 4% SEdm 4% SCat 2% Hert 1% Worc 1% Wolf 0%
10St Edmund Hall 5.7 -1.0 Magd 53% SJoh 23% Ball 12% SEdm 9% SCat 2% Worc 1% Kebl 0% Wolf 0%
11Worcester 5.6 -1.3 SEdm 64% Kebl 15% Worc 8% Magd 7% SJoh 6% Ball 1% NewC 0% Wadh 0%
12New College 5.3 -2.4 Worc 47% Kebl 31% SEdm 17% Magd 3% SJoh 1% NewC 1% Linc 0% Wadh 0%
Division 2
1St John's 6.5 +3.6 Kebl 40% Worc 39% NewC 13% SEdm 3% Wadh 2% Linc 2% Magd 0% SJoh 0%
2Exeter 4.8 -3.4 NewC 45% Wadh 25% Kebl 13% Linc 12% Worc 4% Exet 0% SEdm 0% SJoh 0%
3Keble 5.9 +2.4 Wadh 41% NewC 30% Linc 25% Exet 3% Kebl 1% Worc 0%
4Wadham 5.5 +0.8 Linc 44% Wadh 19% Trin 15% NewC 11% Exet 10% Kebl 0% Lady 0% Jesu 0%
5Lincoln 5.6 +1.3 Trin 57% Exet 31% Linc 4% Wadh 4% Lady 2% Jesu 1% SAnn 0% Kebl 0%
6Jesus 4.8 -2.0 Exet 54% Trin 18% Linc 12% Wadh 9% Jesu 4% Lady 3% Kebl 0% SPet 0%
7L.M.H. 4.9 -2.1 SAnn 99% Jesu 1% Lady 0% Trin 0% SPet 0% Chr2 0% Wadh 0%
8Trinity 5.7 +2.6 Jesu 84% Lady 8% Trin 6% SAnn 1% Linc 0% Exet 0% Wadh 0%
9Queen's 4.6 -2.2 Lady 54% SPet 29% Jesu 5% Bras 4% Trin 3% Chr2 3% Quee 2% SAnn 0%
10Christ Church II 4.8 -2.1 Lady 32% SPet 28% Quee 14% Chr2 13% Bras 7% Jesu 5% Trin 0% Mert 0%
11St Anne's 6.8 +4.0 Quee 47% Bras 23% SPet 17% Chr2 12% Mert 1% Mans 0% Lady 0%
12St Peter's 5.3 +1.5 Quee 36% Chr2 25% Bras 24% SPet 11% Mert 2% Mans 1% Pem2 0%
Division 3
1Mansfield 4.6 -1.5 Chr2 41% Bras 29% SPet 13% Mert 8% Mans 8% Quee 0% Pem2 0%
2Brasenose 5.2 +1.9 Mans 41% Mert 39% Bras 10% Chr2 7% Pem2 2% SPet 1% Quee 0% Some 0%
3Merton 4.8 +0.5 Mans 41% Mert 39% Pem2 14% Bras 4% Some 3% Quee 0% New2 0% SPet 0%
4Pembroke II 4.4 -0.1 Pem2 58% Some 21% Mert 10% Mans 7% New2 3% Lina 0% Bras 0%
5New College II 3.9 -1.2 Some 52% Pem2 23% New2 18% Lina 5% Mert 1% Mans 0% Ori2 0% Corp 0%
6Somerville 4.3 +0.9 New2 52% Lina 26% Some 18% Pem2 2% Corp 2% Ori2 0% Mert 0% Exe2 0%
7Linacre 4.0 -0.2 Ori2 57% Lina 23% New2 11% Corp 5% Some 3% SEH2 1% Exe2 0% Pem2 0%
8Magdalen II 2.8 -3.7 Lina 40% Ori2 32% New2 15% Corp 8% Some 3% Exe2 1% SEH2 0% Mag2 0%
9Corpus Christi 3.7 -0.4 SEH2 46% Corp 36% Ori2 6% Lina 3% Exe2 3% Mag2 2% SHug 1% New2 1%
10Exeter II 3.5 -2.9 Corp 39% SEH2 28% Exe2 9% Mag2 9% SHug 9% Ori2 3% Lina 2% Uni2 1%
11Oriel II 4.6 +3.4 SHug 67% Mag2 18% SEH2 8% Corp 3% Exe2 2% Ori2 1% Uni2 1% SAnt 0%
12University II 3.5 -2.9 Mag2 66% SHug 10% SAnt 7% Corp 6% SEH2 6% Exe2 5% Uni2 1% Ori2 0%
Division 4
1St Edmund Hall II 4.5 +2.9 SAnt 48% Exe2 30% SHug 8% SEH2 6% Uni2 4% Mag2 2% Corp 1% Ori2 0%

[Prev]

Women's Divisions 1-3

Pos Start Crew Score Av.Rise Finishing Crew
Division 1
1Balliol 6.0 -1.3 Pemb 62% Wadh 26% Ball 12% SEdm 0%
2St Edmund Hall 5.6 -2.6 Ball 48% Pemb 29% Wadh 22% SEdm 1% SJoh 0%
3Wadham 6.3 +0.5 Wadh 41% Ball 38% SEdm 11% Pemb 9% Hert 1% SJoh 0% Magd 0%
4Pembroke 6.8 +2.5 SJoh 75% SEdm 17% Wadh 4% Magd 2% Hert 1% Ball 1% OslG 0% Pemb 0%
5Osler-Green 5.3 -2.0 SEdm 68% SJoh 22% Wadh 8% Magd 1% Ball 0% Hert 0% OslG 0% Univ 0%
6Christ Church 4.9 -3.2 Univ 61% Hert 17% OslG 10% Magd 7% SEdm 3% SJoh 2% Wadh 0% Chri 0%
7Magdalen 5.5 -2.3 OslG 77% Univ 17% Hert 3% Magd 1% Chri 0% SJoh 0% Ball 0%
8St John's 6.6 +3.7 Hert 46% Magd 19% Univ 18% OslG 11% Chri 6% SJoh 0% Orie 0%
9Hertford 6.1 +1.0 Chri 75% Hert 17% Magd 3% Univ 3% Mert 1% OslG 0% Orie 0% SJoh 0%
10University 6.7 +3.4 Magd 48% Chri 17% Mert 15% Hert 14% Orie 6% Univ 0% OslG 0% SJoh 0%
11Oriel 5.4 -0.4 Orie 43% Mert 37% Magd 19% Hert 1% OslG 0% Chri 0% NewC 0%
12New College 4.8 -1.1 Orie 51% Mert 46% Chri 2% NewC 1% Worc 0% Wolf 0% Hert 0% Kebl 0%
Division 2
1Somerville 3.9 -3.6 NewC 93% Kebl 4% Mert 1% Lina 1% SCat 1% Orie 0% Worc 0% Wolf 0%
2Merton 5.7 +2.7 Wolf 55% Kebl 29% Lina 7% SCat 4% NewC 3% Some 2% Worc 0% Orie 0%
3Keble 4.6 -0.1 Kebl 47% Wolf 13% SCat 10% Lina 10% Some 9% Worc 8% NewC 2%
4St Catherine's 4.5 -2.1 Worc 59% Some 17% Wolf 14% SCat 4% Lina 3% Kebl 3% SHil 0%
5Linacre 4.7 -1.9 Some 72% Kebl 8% Worc 8% Wolf 4% SCat 4% SAnn 2% Lina 2% SHil 0%
6Wolfson 5.3 +2.8 SAnn 31% SCat 21% Worc 18% Lina 8% Exet 7% Wolf 7% Kebl 5% SHil 2%
7Worcester 5.4 +2.4 SCat 41% SAnn 21% Lina 12% Exet 10% Worc 4% SHil 4% Kebl 4% Wolf 3%
8St Hilda's 4.7 -2.0 Exet 31% Lina 26% SAnn 16% SCat 15% SHil 8% Wolf 2% Worc 2% Quee 0%
9St Anne's 5.2 +1.4 Lina 31% Exet 31% SHil 17% SAnn 16% Mans 4% Wolf 1% Worc 1% Quee 0%
10Exeter 5.3 +1.5 Mans 46% SHil 22% Exet 16% SAnn 8% Quee 8% SCat 0% Wolf 0% Worc 0%
11St Hugh's 3.6 -3.8 SHil 38% Mans 35% Quee 13% SAnn 5% Exet 5% Jesu 4% Linc 0%
12Queen's 4.7 -0.1 Quee 46% Jesu 30% Mans 14% SHil 9% Linc 1% SAnn 0% Exet 0% SHug 0%
Division 3
1Mansfield 5.3 +2.4 Jesu 59% Quee 30% Trin 4% SHug 4% Mans 2% Linc 1% Bras 0% SAnn 0%
2Trinity 4.1 -1.9 Linc 76% SHug 17% Trin 3% Jesu 3% Quee 0% Bras 0%
3Brasenose 3.9 -3.1 SHug 79% Linc 14% Quee 3% Trin 2% Jesu 2% Bras 0% Mans 0%
4Jesus 5.0 +3.2 Trin 82% Corp 7% Linc 7% Jesu 2% Bras 2% Quee 0% SPet 0% SHug 0%
5Lincoln 5.0 +2.7 Lady 71% Corp 14% Trin 8% Bras 6% Linc 0% SPet 0% Jesu 0% ChC2 0%
6St Peter's 3.6 -2.1 Bras 72% Lady 25% Corp 2% Trin 1% SPet 0% Linc 0% Jesu 0% ChC2 0%
7Corpus Christi 4.2 +0.3 Corp 60% Bras 20% SPet 16% Lady 4% Trin 0% ChC2 0% Wad2 0%
8ChC2 3.7 -1.1 SPet 65% ChC2 18% Corp 16% Wad2 0% Lady 0% SJo2 0% Trin 0% SAnt 0%
9L.M.H. 4.9 +3.7 ChC2 52% Wad2 38% SPet 9% SJo2 1% Corp 0% Her2 0% SAnt 0% Lady 0%
10Wol2 2.7 -2.8 Wad2 46% ChC2 29% SJo2 10% SPet 10% SAnt 4% Wol2 1% Lna2 0% Corp 0%
11St Antony's 3.0 -2.0 Her2 75% SJo2 10% Wad2 8% Lna2 3% SAnt 2% Wol2 2% ChC2 0%
12SJo2 3.3 -1.0 Wol2 45% SAnt 30% Her2 20% SJo2 3% Wad2 1% Lna2 1% ChC2 0%
Division 4
1Wadham II 4.0 +3.1 SJo2 38% SAnt 28% Wol2 23% Wad2 6% Her2 4% Lna2 2%

Predicted Gains

[
Prev]

Table shows the crews in decreasing order of predicted Gain in places for the week. Start is the crew's Wednesday starting position. (Eg if a crew has a 50% probability of rising one place and a 50% probability of rising two places, the predicted gain is +1.5.)

Men's Divisions Women's Divisions
CrewGainStartCrewGainStart
1SAnn+4.023 1SJoh+3.7 8
2SJoh+3.613 2Lady+3.733
3Ori2+3.435 3Univ+3.410
4Wolf+3.2 8 4Jesu+3.228
5SEH2+2.937 5Wad2+3.137
6Trin+2.620 6Wolf+2.818
7Kebl+2.415 7Linc+2.729
8Bras+1.926 8Mert+2.714
9SPet+1.524 9Pemb+2.5 4
10Linc+1.31710Worc+2.419
11Pemb+0.9 311Mans+2.425
12Orie+0.9 212Exet+1.522
13Some+0.93013SAnn+1.421
14Wadh+0.81614Hert+1.0 9
15Mert+0.52715Wadh+0.5 3
16Pem2-0.12816Corp+0.331
17Lina-0.23117Quee-0.124
18Corp-0.43318Kebl-0.115
19Magd-0.5 919Orie-0.411
20Univ-0.5 420SJo2-1.036
21Hert-0.7 521NewC-1.112
22SCat-0.8 622ChC2-1.132
23Ball-0.9 723Ball-1.3 1
24SEdm-1.01024Lina-1.917
25New2-1.22925Trin-1.926
26Worc-1.31126SHil-2.020
27Mans-1.52527SAnt-2.035
28Jesu-2.01828OslG-2.0 5
29Lady-2.11929SPet-2.130
30Chr2-2.12230SCat-2.116
31Quee-2.22131Magd-2.3 7
32NewC-2.41232SEdm-2.6 2
33Chri-2.5 133Wol2-2.834
34Uni2-2.93634Bras-3.127
35Exe2-2.93435Chri-3.2 6
36Exet-3.41436Some-3.613
37Mag2-3.73237SHug-3.823


Contributor Statistics

[
Prev]

The following table shows the statistics for each contributor's score compared to the crew scores after the latest SAD iteration. Average error is the average difference, irrespective of sign, between the contributor's score and the current score, while the Bias is the average difference including sign (positive bias means contributor scores are higher than current scores). TT are the scores from last Torpids which provide the initial estimates.

       Pos#   Name   Avg.Error   Bias   No.estimates supplied
         1     NP     0.083     0.083         6
         2     JS     0.100     0.014         7
         3     PC     0.150     0.117        12
         4     JC     0.218     0.073        11
         5     ST     0.311    -0.178        18
         6     BW     0.321    -0.179        34
         7     JJ     0.338    -0.123        13
         8     FM     0.350     0.117         6
         9     SP     0.370    -0.090        10
        10     LG     0.377    -0.115        26
        11     CP     0.408     0.125        12
        12     LL     0.411    -0.078        18
        13     PM     0.414     0.138        29
        14     KS     0.440    -0.120        15
        15     MV     0.479     0.279        14
        16     AD     0.497     0.174        31
        17     TC     0.500     0.050        12
        18     IG     0.567    -0.400         6
        19     TT     0.638    -0.168        74


How to Contribute Information

To supply a score, just decide what sort of standard each crew seems to be, based on the following scale:

Within the SAD program, a difference of 0.4 corresponds to 1.5 lengths quicker over the course, ie a bump is expected at the finish line.

Don't feel obliged to supply an assessment for every crew - just anything you're confident about, even if they are similar to the numbers already listed (it adds weight to the estimate). Also, you can revise your estimates at a later date, in which case I'll remove your earlier estimate for any particular crew. I don't accept scores of your own college.

Normally I will list contributors by their initials (or any other 2-letter acronym you wish to be known by) at the top of this page, but I won't give out any other information on your identity or which crews you have assessed.

Anyone who contributes information will receive the current list of scores for their own college (scores only - contributor codes removed).

New contributors will have a weight +/-0.5 assigned to their estimate. Those who contributed in Torpids 2010 will have their scores weighted by their average error, limited to range 0.2-0.5

Email scores to dudhia@atm.ox.ac.uk


The Gory Details

The SAD program calculates a statistical average of the results of 10000 runs of Bumps. In each run the crew speed is varied by some random amount about the mean speed listed (in brackets), the 1-sigma variation depending on the number, accuracy and time of estimates supplied by contributors. This allows for errors in the assessed speed, but within each `day' of the Bumps, there is an additional term to allow for random variations in a crew's performance from day to day, varying from +/-0.2 for 1st Division crews to +/-0.4 for 3rd Division.
The program allows for overbumps, Sandwich boats, and separate rules for Torpids and Eights, but makes no allowance for crews below those listed.
FORTAN source code.

Contributor estimates are each given a weight of +/-0.5, but those who proved more accurate last Torpids will have those errors used instead. These weights apply for the iteration in which the score was used, but on subsequent iterations all the older scores are reduced in weight (i.e. recent estimates carry more weight).


SAD BaStaRD

(SAD Bayesian Statistical Revised Distribution)

Once racing gets under way, I'll start running SAD BaStaRD (PS Thanks to Tom Fleming for the acronym) to `tune' the actual estimates to values that best fit the results of each day's racing.

The SAD BaStaRD program runs the SAD program and calculates the probability of the actual outcome of each day's racing using the initial scores. It then perturbs each crew's score by +/-0.1, and also the uncertainty in the score by +/-0.1, and recalculates the probability. Whichever perturbation leads to the greatest probability is kept, and the process then repeated. This is continued until, for a given set of scores, all further perturbations result in a lower probability for the actual outcome, and these scores are then taken as the best.

The final result should be scores which best represent each crew's true speed as well as give a means of evaluating the accuracy of each contributor.

If a crew bumps each day, the score that would give the actual outcome with the greatest probability is infinity, so to keep things sensible the original SAD estimates and their uncertainties are used as an a priori constraint.