When people go to ice cream stores with a million flavors, it can sometimes seem impossible to choose. That same thing happens in fantasy football drafts, so I put position analysis at the forefront of the recommendation engine.
The ESPN Draft Assistant analyzes historical fantasy football trends and factors in projections for the upcoming NFL football season to become the best draft companion possible.
The program recommends what position to draft next and gives the user a range of recommended players within that scope. Built for on-the-clock urgency and live-synced updating after every pick.
- Who This Is For
- What It Does
- How The Engine Decides
- How To Use This in a Real Draft
- How to navigate
- The Data Source
- How the Data Is Formatted
- The Math
- Project Structure
- Final Disclaimer and Thanks
This is built for fantasy football players who want confidence and direction in their drafting strategy, rather than solely rankings
Casual players: Casual players want help making picks from a high-level strategy standpoint. Drafting an entire roster without prior knowledge of how to do well is daunting; however, this tool aims to help casual players retain some autonomy in your drafting process while still guiding them towards well-informed picks that will improve their chances of winning it all!
Competitive players: This type of player may already use rankings websites, knows a lot about average draft positions and tiers, but wants to add a tool that factors in the drafting habits of their leaguemates and which position may be the most valuable to take.
Builders/Contributions/Data Scientists: Contributors make this project more meaningful as open-source projects always end up being better. The project is designed to share all of my passion for coding and fantasy football within this project, and I hope advanced users can tune this to their own liking and the option to extend the codebase into something more.
- Modern Dark UI: Professional interface with collapsible sections and intuitive controls
- Visual Status Indicators: Color-coded urgency levels and turn notifications
- Manual Pick System: Comprehensive GUI for testing without ESPN connection
- Debug & Analytics: Detailed breakdowns of scoring calculations and strategic reasoning
There are two decisions every time a user picks a player in fantasy football: What position and who?
-
Position urgency recommendation
The engine will recommend what position is most costly to wait on: in other words, even if there are some very talented players of one position available, it may be more important to take a different position due to the scarcity. -
Top players within that position:
The engine suggests players in the format of a ranked list with how much potential each player possesses, their projected worst possible outcome, and an average outcome for a player with that profile. -
Context Clues
The context clues indicate whether there are time sensitive decisions to make; such as if there are large cliffs in availability. The relevant contextual components being what players have been drafted, what players remain, and how important their position is to the overall success of a fantasy football team.
- Updated recommendations after every pick
- Icons/Emojis for drop-offs
- Color Coding and Status Messages
The core concept centers on two layers of strategy: position urgency and player scoring
- What position is going to get worse the fastest before my next pick?
- Now that I have picked that position, who gives me the best chance given risk, tiers, and ADP (average draft position)?
My goal was to give you autonomy in your final decision after giving the user an array of options. By recommending a position, the end user may then select who they think best fits their needs. Most lists compare players across positions however I attempt to compare players using relative value within their own positions.
For non-technical users, this section may seem confusing. Feel free to contact me if any of this does not make sense. Inside of an IDE (integrated development environment), such as VSCode, open a new terminal. In that terminal, you can then run the below commands.
-
Install dependencies:
pip install -r requirements.txt -
Setup ESPN connection:
python src/setup_mock_draft.py -
Test connection:
python src/tests/test_espn_connection.py -
Run draft assistant (see command below)
-
Run this command in your terminal by opening a terminal
python src/main_gui.py- Configure your league settings Set the scoring format
LEAGUE_ID = 123456789
YOUR_DRAFT_SLOT = 5 # 0-based (5 = pick 6)
TOTAL_TEAMS = 12from config import set_league_type
set_league_type('non_ppr') # or 'half_ppr'- Everything you may need is found in the config.py file
- ESPN Live Sync
⚠️ UNDER CONSTRUCTION
Instructions for this are not needed as of right now, due to API issues But essentially:
Authentication Uses ESPN's cookie-based authentication by going into the developer tools (fn + f12):
espn_s2: Session cookie from browserSWID: User identification cookieleague_id: Extracted from ESPN URL
Data Sources
-
Direct Draft Picks:
league.draftendpoint -
Team Rosters:
league.teams[].roster(most reliable during live draft) -
Draft Recap: ESPN's
mDraftDetailendpoint (best for mock drafts) -
League Settings: Fallback for draft status
-
During the draft Recommendations update automatically after each pick. You must enter your own picks in the GUI. Use the "➕ Add Pick" to manually mark picks
-
Make your pick Review recommendations, check position analysis, and select your player
- By Position: Click QB/RB/WR/TE buttons to see top 10 players at that position
- Overall Analysis: Shows position urgency breakdown for all positions
- Urgency scores (higher = more urgent)
- Drafted count vs. position limits
- Roster needs (critical/high/medium/low priority)
- 🟢 Tier 1: Elite players (highest value)
- 🟡 Tier 2: Strong players
- 🟠 Tier 3: Solid players
- 🔴 Tier 4+: Lower-tier players
- 📉 Cliff: Large dropoff detected (positional scarcity warning)
- ⏰ Early QB: QB recommended too early (penalty applied)
For the majority of the math in this codebase, the data source is the genius behind the calculations.
- The program utilizes this data source from FFAnalytics. The projections are not just a singular number, but have variance, cutoff, VOR, and a range of numbers over simulations. Their methodology optimizes your draft strategy by allowing you to make statistics based decisions and understand the probability of outcomes.
Due to the wide range of data, I was able to provide lots of advanced analytics.
- Ceiling/Floor Analysis: The top 10 players sorted by their ceiling with detailed breakdown on their range of potential outcomes
- Uncertainty Scoring: An uncertainty score provides depth into the risk assessment on each player which is based on player volatility and projection uncertainty
- ADP (Average Draft Position) Value Analysis: Reach cost calculations and value opportunity identification based on where certain players are selected in drafts, on average
- Tier-Based Visualization: Color-coded tiers (🟢🟡🟠🔴) with cliff and indicators if it is too early to take certain positions
- points: baseline projection
- floor: the projected worse case scenario of total points a player may accumulate over a season (excluding injury)
- ceiling: the projected best case scenario of total points a player may accumulate over a season (excluding injury)
- uncertainty: a penalty based on the variability this player may experience
- tier: The player quality tier, the lower the number is, the more elite of a status that player tends to maintain.
- ADP (average draft position): The market consensus draft position of that player, aggregated across multiple platforms and mostly calculated from mock drafts from real users.
- dropoff: the value drop to the next player at that position; used to detect urgency and cliffs.
- sd_pts: the standard deviation of points utilized as a measure of volatility
- cliff_bonus: The bonus for players with large dropoffs meant to protect users from positional scarcity.
- adp_value: A bonus for players that go later than expected as it may be beneficial to go with public consensus at times
- composite_score: An overall value rankings for relative position comparisons
composite_score =
vor_score + # 70% weight
tier_bonus * 0.15 + # 15% weight
cliff_bonus * 0.10 + # 10% weight
adp_value * 0.05 + # 5% weight
risk_tilt - # Round-dependent
reach_cost # Penalty- VOR stands for value over replacement, essentially acting as an overview on how valuable a player is projected to be compared to the average replacement. The final score weights the VOR very heavily while factoring in other variables:
score = (
vor * 0.75 + # Base VOR (75%)
ceiling_vor * 0.10 + # Ceiling VOR (10%)
floor_vor * 0.10 + # Floor VOR (10%)
vor * (tier_multiplier - 1.0) + # Tier bonus (scales with VOR)
dropoff_bonus + # Cliff detection
adp_value + # Market value
- uncertainty_penalty + # Risk management
opportunity_cost + # Opportunity cost analysis
positional_adjustment + # Roster balance
round_adjustment # Floor/ceiling preference
)Here comes the methodology behind the recommendation engine. There are not a lot of advanced techniques behind this, and I would love to receive feedback on improvements or additions with machine learning to optimize performance.
Urgency determines how much value you lose at a position if you wait until a further pick to choose this position. The system employed a within-position normalization process to compare positions fairly. Normalization levels the comparison field so decisions can be comparable while otherwise harder to do so (attempting to compare apples to oranges). However, the engine attempts to favor RBs and WRs more as they tend to be the main reason people win fantasy championships. After normalizing the positional VOR distribution, the engine measures how the best available player degrades relative to that position.
- Breakdown
- 📈 Opportunity Cost (60% of the weight)
- Measures how important it is to take that player now compared to their replacements at a further point in time. The logic here utilizes percentile ranks within that position (e.g., if the 10th percentile RB drops to the 22nd percentile), that registers as a 12 point loss in relative value. By capturing the cost of waiting, the present decision is weighted properly.
⚠️ Cliff Pressure (25% weight)- Adds urgency when there is a significant dropoff to the next tier of player. If a decision is urgent, then it must be made now. This normalizes dropoff values, but caps the maximum when the change is over 25 VOR points, to protect the user from positional scarcity. A large cliff indicates that waiting leaves you with weak options
- 📋 Roster Need (15% weight)
- If the urgency for a position is critical, but you don't need them nor the depth, then this roster need calculation accounts for that. This prioritizes filling starter requirements and proper FLEX depth (RB/WR/TE) based on your scoring format.
- 🚫 Early QB Penalty
- This constant reduction encourages users not to draft QBs too early if the advantage is not significant as empirical evidence points to the strategy of drafting a QB in a 1-QB league can be detrimental to your chances.
- 📈 Opportunity Cost (60% of the weight)
Player scoring ranks candidates within selected filters using their value, tiers, cliffs, ADP, and round-specific risk preferences. Once a position is selected, players are then scored within that position only. This avoids cross-position VOR bias and lets you compare players on a level playing field. The composite score from above includes the formula for this.
| Round | Formula | Strategy |
|---|---|---|
| Early (1-5) | (floor - vor) * 0.2 - uncertainty * 0.1 | Prefer floor and consistency |
| Mid (6-8) | ((floor + ceiling) / 2 - vor) * 0.1 - uncertainty * 0.05 | Balanced approach |
| Late (9+) | (ceiling - vor) * 0.2 - uncertainty * 0.05 | Prefer ceiling and upside |
Early rounds favor safe, high-floor players. Mid rounds balance floor and ceiling. Late rounds favor high-ceiling players. Therefore, any uncertainty is penalized more in early rounds.
- Reach Costs
These are meant to be a guardrail rather than a strict rule to adhere to. This discourages reaching too far above a player's ADP but some exceptions do exist such as:
- Large dropoffs
- Critical roster needs
- Elite tier players
- ADP is the aggregate consensus on where players are commonly selected in fantasy football drafts and can change based on the league format.
- I configured a dynamic ADP reach penalty to ensure the user is still drafting closer to market trends based on an interesting article where draft selection is like picking stocks.
- Early Picks (1-36): 1.5x penalty multiplier, stricter thresholds
- Mid Picks (37-84): Standard penalties
- Late Picks (85+): 0.7x penalty multiplier, more lenient thresholds
- Non-PPR: Prioritizes RBs, penalizes WRs slightly
- Half-PPR: Boosts WRs, adjusts TE values
- Position weights scale with actual VOR values
VOR is value over replacement. This value essentially indicates that if this player was replaced by the average production of a player with the same position.
- The new methodology is to not compare VOR (value over replacement) between different positions because running backs on average, possess a higher mean VOR. So, now we analyze intra-positionally which players might be drafted (based on composite ADP) in between your next pick. This produces an urgency value, what position will you benefit from the most given your roster composition and the available players. This factors in the standard deviations from the FFAnalysis data set.
espn-draft-tracker/
├── src/
│ ├── main_gui.py # GUI application
│ ├── main.py # Console version
│ ├── config.py # Configuration
│ ├── recommendation_engine.py # VOR-based engine
│ ├── data_loader.py # Data loading
│ ├── draft_analyzer.py # Draft analysis
│ ├── utils/ # Utility scripts
│ └── tests/ # Test suite
├── projections_*.csv # Data files
└── requirements.txt # Dependencies
First and foremost, projections are not guarantees. This tool depends on statistical projections based on historical data and simulations. However, projections are merely a reflection of probabilities, not reliabilities. They are the expected outcomes and not the certainties. Player performance will vary for many reasons, this tool is meant to support decision making, not replace it. Use recommendations as one input among many. Also, currently there are no injury projections therefore it is hard to factor that into the engine's decision making even though it is a big part of fantasy football. The algorithm is also limited in its current state as it is meant for single quarterback leagues and only supports half PPR and non-PPR. Hopefully, with others contributions, this tool can expand to superflex, auction, keeper, dynasty, full PPR, and more league options. This engine also depends only on FFAnalytics for projection data and if they change their data format or stop providing data, this system will need lots of adjustment. There is no guarantee in results, but it is very fun to try.
Created by Harrison Dolgoff
Thank you to FFAnalytics for the open source data set and ESPNApi for connecting to live ESPN draft data