Last year, I used machine learning to rank the running backs. The results were mixed: I was right about Dalvin Cook, David Johnson, and Jordan Howard, but wrong about LeSean McCoy, Royce Freeman, and Melvin Gordon. Overall, I was right on 9 predictions, and wrong on 9 predictions, and ranked below average compared with the experts.

I’ve made some improvements for this season: my model isn’t as needlessly complex, and my predictions on everyone’s carries and stats are more realistic. My model still made some bold calls, but I only see one outlandish predictions.

The Rankings:

Sorry for the formatting.

Player	Carries	Rush Yards	Rush TDs	Rec	Rec Yards	Rec TDs	Points	ADP
1. Saquon Barkley	255	1149	9	76	616	4	251	1
2. Alvin Kamara	191	890	9	81	695	4	235	2
3. Nick Chubb	254	1194	9	43	326	2	219	7
4. Christian McCaffrey	184	907	6	82	681	4	218	3
5. Ezekiel Elliott	237	1058	7	54	428	2	204	5
6. David Johnson	230	937	7	52	490	3	201	4
7. Joe Mixon	246	1104	8	45	337	1	198	10
8. Todd Gurley	200	909	8	47	413	3	197	9
9. Le’Veon Bell	247	1035	7	49	404	2	195	6
10. James Conner	210	923	8	46	392	2	192	8
11. Melvin Gordon	185	858	7	44	370	2	181	14
12. Dalvin Cook	212	951	6	49	376	2	179	11
13. Leonard Fournette	242	966	7	36	290	1	177	13
14. Derrick Henry	227	1037	9	16	123	0	171	21
15. Marlon Mack	217	960	8	27	209	1	169	16
16. Damien Williams	168	755	7	44	352	3	168	12
17. Kerryon Johnson	194	907	6	43	318	2	165	15
18. Aaron Jones	183	869	7	34	261	1	161	17
19. David Montgomery	200	856	6	32	261	1	157	20
20. Devonta Freeman	184	792	6	34	271	1	151	18
21. Josh Jacobs	193	838	5	35	284	1	151	19
22. Chris Carson	200	894	7	23	183	1	151	22
23. Philip Lindsay	176	824	6	33	239	1	147	25
24. Kenyan Drake	142	655	4	48	391	2	145	29
25. Mark Ingram	209	856	6	28	209	1	144	23
26. Lamar Miller	194	846	4	27	210	1	138	27
27. Sony Michel	193	840	7	15	110	0	137	24
28. Tevin Coleman	153	708	5	30	251	2	136	26
29. Latavius Murray	182	726	7	24	171	1	135	32
30. James White	69	314	2	66	558	4	123	31
31. Miles Sanders	156	634	4	31	251	1	122	28
32. Austin Ekeler	108	522	3	39	351	2	120	34
33. Derrius Guice	154	655	5	24	206	1	120	30
34. Rashaad Penny	152	671	4	25	198	1	119	33
35. Tarik Cohen	74	345	3	54	482	3	115	35
36. Royce Freeman	161	652	5	21	156	1	113	38
37. Darrell Henderson	136	579	4	28	227	1	111	37
38. Kalen Ballage	143	610	7	23	188	1	111	45
39. Jordan Howard	161	662	5	15	112	0	110	36
40. LeSean McCoy	155	620	3	31	236	1	109	41

The Rest:

#	Player	Points	ADP
41	Ronald Jones	105	40
42	Carlos Hyde	101	42
43	Alexander Mattison	98	48
44	Jaylen Samuels	95	46
45	Devin Singletary	92	47
46	Peyton Barber	91	43
47	Justin Jackson	91	52
48	Kareem Hunt	91	39
49	Adrian Peterson	91	44
50	Damien Harris	87	50
51	Duke Johnson	87	51
52	Tony Pollard	86	56
53	Nyheim Hines	84	60
54	Matt Breida	83	49
55	D’Onta Foreman	80	59
56	Darwin Thompson	80	55
57	Justice Hill	79	53
58	Mike Davis	78	58
59	Dion Lewis	76	56
60	Ito Smith	74	59
61	Jerick McKinnon	70	54
62	C.J. Anderson	66	61
63	Malcolm Brown	65	62
64	Jalen Richard	57	—
65	Jamaal Williams	54	—
66	Theo Riddick	53	—
67	Giovani Bernard	51	—
68	Frank Gore	46	—
69	Alfred Blue	45	—
70	Gus Edwards	45	—
71	Dontrell Hilliard	45	—
72	Chris Thompson	45	—
73	Wayne Gallman	44	—
74	Ty Montgomery	43	—
75	Chase Edmonds	42	—
76	Mark Walton	39	—
77	Ryquell Armstead	39	—
78	Rod Smith	39	—
79	Benny Snell	38	—
80	Cameron Artis-Payne	37	—
81	Bilal Powell	36	—
82	Qadree Ollison	35	—
83	Dexter Williams	34	—
84	Jordan Scarlett	32	—
85	Alfred Morris	32	—
86	Andre Ellington	30	—
87	Zach Zenner	28	—
88	Doug Martin	27	—
89	Trayveon Williams	26	—
90	Ameer Abdullah	24	—
91	J.D. McKissic	23	—
92	Zach Line	22	—
93	T.J. Logan	17	—
94	David Fluellen	17	—
95\	Josh Ferguson	5	—
96	Darren Sproles	-4	—

Bold Picks:

Sleepers and Undervalued:

Nick Chubb: Rank: 3, ADP: 7
Derrick Henry: Rank: 14, ADP: 21
Joe Mixon: Rank: 7, ADP: 10
Melvin Gordon: Rank: 11, ADP: 14
Kenyan Drake: Rank: 24, ADP: 29
Kalen Ballage: Rank: 38, ADP: 45
Nyheim Hines: Rank: 53, ADP: 60

Busts and Overvalued:

Le’Veon Bell: Rank: 9, ADP: 6
David Johnson: Rank: 6, ADP: 4
Damien Williams: Rank: 16, ADP: 12
James Conner: Rank: 10, ADP: 8
Kareem Hunt: Rank: 48, ADP: 39
Jerick MicKinnon: Rank: 61, ADP: 54

Last year, I trained a machine learning model to rank the running backs based on how it expected them to do during the 2018 season. The resulting rankings were different from the consensus, to put it mildly. When I posted my rankings, the results were mixed. Some were interested and wanted to learn more, some thought that my rankings were unrealistic, and others accurately pointed out flaws with my methodology.

Now that the 2018 season is in the books, it’s time to evaluate how my model did!

First, a look at the boldest predictions:

Dalvin Cook: Model: RB27, ADP: RB10, Actual Rank: 31st
“What’s up with Dalvin Cook? “F” for efficiency? I get the injury concern but…that makes no sense”
“I’m not a huge Dalvin Cook truther or anything, but come on man, you know that his ranking is absurdly low here unless the model is predicting him to be hurt again”
When I posted these predictions, this was the one I got the most criticism on. Cook was an exciting running back who had a very good start to the 2017 season before getting injured. He was supposed to be healthy in 2018 and everyone expected him to bounce back. However, I thought he would finish in 27th and finish one spot below his backup, Latavius Murray. Actually, Cook finished in 31st and finished one spot above Murray.
Result: CORRECT

David Johnson: Model: RB7, ADP: RB3, Actual Rank: 10th
“Did DJs injury influence the results? He seems crazy low at 7.”
Last year, David Johnson was one of four running backs considered to be in the top tier. Me ranking him all the way down at 7th was pretty controversial. However, David Johnson was weighed down by an awful Cardinals offense and he ended up finishing 10th, meaning that my prediction was actually too optimistic.
Result: CORRECT

Latavius Murray: Model: RB26, ADP: RB44, Actual Rank: 32nd
“Latavius Murray over Kerryon Johnson huh”
Murray being ranked so high was a residual effect of my model expecting Dalvin Cook to bust. Dalvin Cook did bust and Murray did get more of an opportunity, but he didn’t really take advantage of it. He did do much better than most people thought, though.
Result: Slightly Correct

LeSean McCoy: Model: RB10. ADP: RB17, Actual Rank: 40th
“If McCoy finishes above Gordon I’ll shit a brick”
Yeah, this pick was pretty bad. McCoy was getting older and stuck on the Bills offense, but he looked like he was going to get a lot of carries. Instead, he only got 161 carries and came in 40th in points.
Result: INCORRECT

Royce Freeman: Model: RB11, ADP: RB18, Actual Rank: 47th
“there is 0 way royce freeman sees 50 more touches than melvin gordon.”
This was by far the worst pick my model made. Royce Freeman looked like an exciting rookie running back. People were already optimistic on him, but my model took that to the extreme, expecting him to close to being a top 10 RB in his first season. A Broncos rookie running back did almost make the top 10, but it was Philip Lindsay, not Royce Freeman. To be fair, nobody saw Lindsay coming, and I didn’t even make a prediction for him. Maybe if I did, Royce would have been ranked lower. But still, Royce didn’t even look impressive in the playing time that he did get, and ranking him 11th is bad.
Result: INCORRECT

Jay Ajayi: Model: RB14, ADP: RB22, Actual Rank: 78th
“In what universe is Jay AJayi getting more carries than Melvin Gordon?”
After a solid 2016, Ajayi had a down year in 2017. My model expected him to bounce back, but that didn’t happen. He had a very good first game, but was then put on injured reserve after only three more games.
Result: INCORRECT

Melvin Gordon: Model: RB12, ADP: RB8, Actual Rank: 6th
“If McCoy finishes above Gordon I’ll shit a brick”
“Melvin Gordon at 12 is all I need to know”
When I posted my rankings on reddit, most of the comment section was criticizing me for ranking Melvin Gordon too low. I argued that I thought Austin Ekeler would be more involved this year. While that turned out to be true, my prediction about Gordon was way off, as he finished in 6th.
RESULT: INCORRECT

Jordan Howard: Model: RB17, ADP: RB12, Actual Rank: 20th
“The disrespect for Jordan Howard is tragic…especially since this is a standard league. Ranked below McCoy, Mixon, Royce Freeman, Ajayi and Henry? Come on now…”
I got plenty of criticism for putting Howard so low, especially with these rankings being for standard scoring. But after two years of being a top 10 running back, Howard finished in 20th, performing even worse than my model expected. In the offseason, he was traded for only a fifth round pick.
Result: CORRECT

Ezekiel Elliott: Model: RB2, ADP: RB4, Actual Rank: 5th
My model was very optimistic on Zeke, predicting that he would be the second best running back. While he didn’t quite live up to that, he still had a great season. An owner who selected Zeke would probably be pretty happy, given that the other running backs in that tier were Le’Veon Bell and David Johnson.
Result: Slightly Correct

Kareem Hunt: Model: RB5, ADP: RB9, Actual Rank: 8th
After the first eleven weeks of the season, Hunt had scored the third most points among running backs, and this prediction looked very good. Then after a video was released of him kicking a woman, he was released. Overall, he finished as the RB8, one spot above his ADP.
Result: Slightly Correct

Duke Johnson: Model: RB32, ADP: RB50, Actual Rank: 48th
My model expected that Duke and Carlos Hyde would be the surprising Browns running backs, and that Nick Chubb would be a disappointment. That turned out to be backwards, and Duke struggled to make an impact in a crowded Browns backfield.
Result: INCORRECT

Alvin Kamara: Model: RB4, ADP: RB6, Actual Rank: 4th
Although some people did have Kamara at the top of the second tier of RBs, most people predicted regression. And while that did happen, that was more than offset by an increased number of carries and 18 total touchdowns. Overall, he was the fourth highest scoring running back, which was exactly what my model predicted.
Result: CORRECT

Rex Burkhead: Model: Unranked, ADP: RB28, Actual Rank: 80th
Burkhead got injured during the season. But when he came back, he was stuck in a crowded backfield and never broke 40 rushing yards.
Result: CORRECT

Isaiah Crowell: Model: RB24, ADP: RB34, Actual Rank: 29th
Crowell had a better year than most people expected, but didn’t do as good as my model thought. He finished at 29th, exactly in the middle between my prediction and his ADP.
Result: Slightly Correct

Christian McCaffrey: Model: RB8, ADP: RB11, Actual Rank: 3rd
My model was pretty optimistic. In fact, I didn’t see any rankings that had CMC higher than 8th. But McCaffrey surpassed even that prediction, totaling almost 2000 yards from scrimmage and finishing as the 3rd highest scoring running back.
Result: CORRECT

Jamaal Williams: Model: RB37, ADP: RB27, Actual Rank: 46th
With Aaron Jones suspended for two games, Williams was the highest ranked Packers RB. But with the state of the Packers backfield unclear, as well as Williams’s struggles in 2017, my model was low on him. That turned out to be correct, as Williams got only 121 carries even with Ty Montgomery being traded midseason.
Result: CORRECT

Ronald Jones: Model: RB29, ADP: RB39, Actual Rank: 100th
Ronald Jones had a pretty awful year, and made this pick look very bad. He was a healthy scratch for the first three weeks and then had only 1.9 yards per carry when he actually got on the field.
Result: INCORRECT

Chris Carson: Model: RB38, ADP: RB29, Actual Rank: 14th
This prediction was a definite miss. My model had Carson four spots behind his projected backup, Rashaad Penny. In reality, Carson seized the starting job and finished fifth in the NFL in rushing yards.
Result: INCORRECT

Peyton Barber: Model: Unranked, ADP: RB31, Actual Rank: 26th
Barber slightly surpassed expectations, but my model didn’t even list him. This was probably a side effect of having Ronald Jones so high.
Result: INCORRECT

Carlos Hyde: Model: RB20, ADP: RB25, Actual Rank: 44th
Hyde actually had a pretty good start to the season. After six weeks, he had the 12th most points, but was then traded to the Jaguars to make room for Nick Chubb. Hyde struggled with his new team, with only 58 carries and 3.3 yards per carry, and finished in 44th.
Result: INCORRECT

Sony Michel: Model: RB28, ADP: RB35, Actual Rank: 25th
I was one of the highest on Michel, and he turned in a good season, getting over 200 carries and rushing for almost 1000 yards.
Result: CORRECT

Dion Lewis: Model: RB35, ADP: RB30, Actual Rank: 36th
After going from the Patriots to the Titans and splitting time with Derrick Henry, Lewis wasn’t expected to repeat his 2017 season. My model thought that he wasn’t downgraded enough, though, and ranked him at 35th. Lewis ended up finishing at 36th.
Result: CORRECT

Tally:
Correct: 9
Incorrect: 9
Slightly Correct: 4

Overall, the results are mixed. My model had some very big wins (Dalvin Cook, David Johnson, Jordan Howard) but also some big misses (Royce Freeman, Melvin Gordon, LeSean McCoy). Interestingly, my model seemed to be more accurate at predicting busts than predicting sleepers.

Now that we’ve looked at the individual bold predictions it made, let’s look at how the model did in general. I scoured the internet and found twenty nine other rankings made for the 2018 season. I went through all of them and calculated both their ECR (lower is better) and their weighted points (calculated by multiplying their #1 RB’s actual points by 25, their #2 RB’s actual points by 24 and so on for their top 20 RBs). Then I did the same thing for my model and compared the results.

Expert	ECR	Weighted Points
Scout Fantasy Sports – Dr Roto	2942.5	74171.7
Yahoo – Pianowski	3031.9	73959
TheFootballGirl	3036.7	73243.4
Scout Fantasy Sports – Atkins	3076.1	75640.2
Scout Fantasy Sports – Brandon	3082.9	75253.6
FantasySixPack – Bond	3114.5	73139.2
Yahoo – Behrens	3119.3	72471
Razzball	3147.1	73759.5
FantasySixPack – Savill	3149	73847.2
WalterFootball	3194.4	70000.6
Scout Fantasy Sports – Childs	3200.2	67176.1
ESPN	3236.9	74029.6
FantasySixPack – Lott	3250.6	72456.6
Scout Fantasy Sports – Ronis	3252.8	73284.5
Yahoo – Del Don	3253.4	69348.5
FFToday	3336.1	67260.7
Yahoo – Evans	3371.2	69082.8
5th Down Fantasy – Sablich	3401.3	73186.4
5th Down Fantasy	3417.4	70610.8
SportingNews	3425.4	71126.9
FantasyShed	3440.5	70217.4
Yahoo – Loza	3446.1	71419.3
CBS – Dave Richard	3479.6	72015.3
Simple Model	3651.6	70556.8
FantasySixPack – Hamrick	3670.9	70676.4
Fantasy Football Index	3734.8	65426.2
RotoWorld	3746.2	69722.8
RotoProfessor	3747.2	71809.1
Complex Model	3833.3	70478.1
PitcherList	3950.4	66787.5
FakePigskin	4059.7	66612.5

Overall, these results aren’t great. My simple model came in 24th out of 31 for ECR and my complex model came in 29th. The results for weighted points are better: the simple model came in 20th and the complex model came in 21st. These results are slightly misleading: I posted my results right before Jerick McKinnon got injured, and most of the other rankings didn’t include him.

Even if I take out the McKinnon prediction though, the results aren’t great. I was hoping that my model would blow the experts away and that didn’t happen. However, the results aren’t terrible for a first attempt. I’m looking forward to fixing the flaws in my model and giving it a try for 2019.

Verdict: The Machine needs to go back to school.

The most important part of a fantasy football season is the draft. Everyone starts off with high hopes and a blank roster. Everyone has the same information and, with the exception of the first round, the opportunity to get any player they want.

Seasons are won and lost at the draft. Picking up a guy in the 7th round who goes on to finish as an RB1, or nailing a late round sleeper pick could propel you into the playoffs. On the other hand, if you miss on a first or second round pick, you’re starting off at a big disadvantage.

What I’ve said isn’t new. Each year, dozens of cheat sheets, draft kits and rankings come out.

But most of these rankings look the same.

Sure, some rankings might have a guy going a couple picks, or even half a round earlier than another set of rankings. At the end of the draft, the difference might be a couple of rounds. But for the most part, the rankings are very similar.  Everyone agrees that there are four running backs that stand out from the rest. And after that is another tier of about a half dozen guys who will go in the back of the first and the early second. After that are rookies, guys who have struggled, and backs dealing with competition.
For the most part, these rankings are reasonable. Todd Gurley, LeVeon Bell, David Johnson and Ezekiel Elliott are all very talented running backs, and the experts who rank them are very knowledgable about fantasy football.

But they’re human, and all humans have biases. Young running backs with upside to hit 1000 yards get taken ahead of older running backs that already do. Guys with a lot of name value get taken ahead of where they should go. If you drafted a guy last year who busted, chances are you’re going to hesitate before drafting him this year.

But what if these running backs could be predicted by machine learning? The computer doesn’t care what name a player has, and it won’t overreact to how a player performed on their fantasy team last year, because it didn’t have a fantasy team. Instead, it will make impartial predictions based on production, opportunity, age, injury history, combine results, and the team around them.

I gathered data from the 2008 to the 2017 seasons and fed them in. I used six models: Linear Regression, Logistic Regression, K Nearest Neighbors, Random Forest, Decision Trees, and Linear Discriminant Analysis. I then had two different ways of turning the results from those models into rankings: A) averaging the rankings from the models, and B) feeding the results into a second linear regression model. When I tested these methods, both models performed about the same, so the results from the simpler model are my official projections.

With the details out of the way, here are my rankings for 2018 running backs in standard leagues! The first column is the rankings from the simpler model. The second column is the rankings from the more complex model. The next two columns are each player’s predicted carries, and a letter grade representing how efficient I expect them to be at turning carries into points. (I also aggregated my rankings with ADPs found at Fantasy Football Calculator, so if the rankings don’t match up, that’s probably why).

#	Player	Simple Ranking	Complex Ranking	Carries	Efficiency
1	Todd Gurley	1	1	244	A+
2	Ezekiel Elliott	2	3	226	A-
3	Le’Veon Bell	3	4	203	A+
4	Alvin Kamara	4	7	193	A
5	Kareem Hunt	5	5	235	B
6	Saquon Barkley	6	6	197	C+
7	David Johnson	7	2	182	A-
8	Christian McCaffrey	8	8	153	A
9	Leonard Fournette	9	11	206	B-
10	LeSean McCoy	10	15	275	C
11	Royce Freeman	11	18	231	B+
12	Melvin Gordon	12	12	184	B+
13	Joe Mixon	13	19	218	A
14	Jay Ajayi	14	23	201	B
15	Derrick Henry	15	13	182	B+
16	Devonta Freeman	16	9	180	B
17	Jordan Howard	17	10	208	C
18	Alex Collins	18	14	193	A-
19	Lamar Miller	19	22	195	B-
20	Carlos Hyde	20	16	194	C+
21	Jerick McKinnon	21	21	206	B
22	Kenyan Drake	22	25	132	A+
23	Marshawn Lynch	23	26	214	C
24	Isaiah Crowell	24	33	166	A
25	Mark Ingram	25	20	122	A-
26	Latavius Murray	26	31	177	B-
27	Dalvin Cook	27	24	146	F
28	Sony Michel	28	29	225	D-
29	Ronald Jones	29	40	215	C+
30	Kerryon Johnson	30	17	165	D-
31	Tevin Coleman	31	30	141	D+
32	Duke Johnson	32	36	118	A+
33	Tarik Cohen	33	39	131	A
34	Rashaad Penny	34	27	179	D+
35	Dion Lewis	35	28	133	B-
36	Marlon Mack	36	43	157	C-
37	Jamaal Williams	37	34	100	C+
38	Chris Carson	38	32	130	D+
39	LeGarrette Blount	39	35	159	D+
40	C.J. Anderson	40	46	123	C-