Weighted Averages in Fantasy Projections: How Past Performance Is Scaled
Not all games are created equal — and projection models know it. When a model builds a statistical forecast for a fantasy player, it doesn't simply average every game from the past two seasons and call it a day. It applies weights, assigning more influence to recent performance, more relevant sample sizes, and conditions that more closely resemble what's coming next. That selective scaling is the mechanism behind weighted averages in fantasy projections, and understanding how it functions explains why two systems can look at the same player and arrive at meaningfully different numbers.
Definition and scope
A weighted average is a mathematical technique where individual data points contribute to a final figure in proportion to an assigned weight rather than equally. In standard arithmetic, five games producing 8, 12, 6, 14, and 10 fantasy points yield a simple mean of 10.0. In a weighted framework, those same five games might produce a result closer to 11.5 if the two most recent performances carry three times the influence of the earliest game.
The scope of this technique inside fantasy projection systems is broad. Weights are applied across time (recent weeks vs. earlier weeks), sample quality (full-game snaps vs. injury-shortened appearances), opponent difficulty, and situational context like weather or dome/outdoor splits. The glossary of projection terms covers the full vocabulary surrounding these inputs, but weighted averaging is the arithmetic layer that actually operationalizes their relative importance.
How it works
The mechanics follow a consistent structure, regardless of the projection system.
- Raw data collection — historical game logs, season totals, or rolling windows are assembled. Most systems use somewhere between 8 and 32 games as a baseline window.
- Weight assignment — each observation receives a coefficient. A common decay function is exponential: game n weeks ago carries a weight of λⁿ, where λ is a decay constant between 0.7 and 0.95. At λ = 0.85, a game from 6 weeks ago retains roughly 38% of the influence of last week's game (0.85⁶ ≈ 0.377).
- Weighted sum calculation — each data point is multiplied by its weight, and those products are summed.
- Normalization — the sum is divided by the total of all weights, not the count of observations. This is the step that distinguishes a weighted average from a weighted sum.
- Contextual overlays — the resulting figure is then adjusted for opponent, usage rate, and scoring format before output. Those adjustment layers are described in detail under usage-rate adjustments in projections.
The decay constant λ is a genuine design decision with real consequences. Lower values (closer to 0.7) make the model highly reactive to recent performance — useful during volatile stretches of a season, but prone to overreaction to single-game outliers. Higher values (closer to 0.95) produce smoother, more stable projections but can lag when a player's role genuinely changes, as explored in in-season vs. preseason projections.
Common scenarios
Returning from injury — If a running back missed four games and returned to 60% snap counts in weeks five and six before reaching full workload in weeks seven and eight, a naive recent-game weighting will undercount his true role. Well-designed systems either exclude injury-limited games from the weighted window or apply a usage-normalized weight based on snaps per game rather than raw point totals. The connection between snap counts and projection accuracy is covered under snap count and target share data.
Breakout pace vs. regression risk — A wide receiver averaging 22 fantasy points per game over his last three games looks extraordinary until regression to the mean in fantasy mechanics are applied. A model using a wide window of 20+ games will pull that figure back toward a longer-run baseline, which is often the statistically correct call — even when it feels pessimistic.
Comparing recent form to preseason projections — Preseason systems weight career-level production and aging curves heavily, since there are no current-season data points. By week ten of an NFL season, that same system may weight the ten most recent games at 70% of total projection influence, nearly inverting the original balance. Projection models explained describes how these architectural shifts are managed across a full season.
Decision boundaries
The critical decision in any weighted-average system is where to draw the boundary between "informative data" and "noise too old to matter." Three comparison cases illustrate how that line is drawn differently across contexts.
Short window (4–6 games) vs. long window (16–24 games): Short windows are appropriate for volatile positions — running backs with fluctuating usage, relievers in MLB, or players entering new offensive systems. Long windows suit quarterbacks and starting pitchers, whose underlying skill signals are more stable across time. Sample size and projection reliability addresses the statistical tradeoff directly.
Equal decay vs. hard cutoff: Some systems use a hard cutoff — games beyond week 12 of the previous season simply receive a weight of zero. Others use continuous exponential decay with no hard boundary. The decay model is generally more statistically defensible; the hard cutoff is simpler to explain and audit.
Position-specific calibration: No single λ value performs optimally across all positions. Research published by analysts at Rotoviz and Sports Info Solutions has consistently shown that target-share-driven receivers tolerate wider windows than committee backfields, where role changes can render three-week-old data nearly irrelevant overnight. The scoring format impact on projections page connects these weighting decisions to the downstream question of PPR vs. standard scoring output.
The Fantasy Projection Lab home applies weighted averaging as a foundational layer beneath every positional model, with position-specific decay constants calibrated through backtesting projection accuracy against multiple seasons of historical outcomes.