1. The "Dormant Event" Paradox
In competitive swimming, an athlete's race profile is fundamentally asynchronous. Because competition schedules and training cycles emphasize specific strokes or distances across different phases of the season, swimmers rarely race their entire 18-event repertoire regularly.
A distance freestyler may race the 100m, 200m, and 400m Freestyle every month to tune speed and race tempo, but contest the grueling 800m or 1500m Freestyle only once every 12 to 24 months. Similarly, an Individual Medley specialist often spends an entire winter racing single-stroke 100m events, leaving their 400m IM un-raced for an entire year.
This creates the Dormant Event Paradox:
- Aerobic and neuromuscular expansion occurs across active events: The swimmer's overall engine, stroke volume, lactate buffering, and velocity at critical thresholds have advanced significantly over the preceding 12 months.
- The seed record in the dormant event remains frozen in time: On paper, the athlete's personal best (PB) in the dormant event reflects their physiological capacity from one or two years prior.
- Traditional longitudinal models fail: Point-to-point time predictors project forward strictly from the ancient seed swim along generic population aging curves, remaining completely blind to the fact that the athlete has dropped two seconds in their 200m events over the intervening period.
- Pure cross-event imputation models discard valuable signal: Conversely, zero-shot models (such as our foundational Project M11 engine) infer the target event exclusively from current swims, strictly ignoring the athlete's prior PB and erasing their unique, hard-earned biomechanical affinity for that specific stroke and distance.
Neither isolated longitudinal extrapolation ($t_1 \to t_2$) nor blind cross-event imputation at $t_2$ captures the full picture. True performance requires projecting historical stroke affinity forward along the path of verified engine growth.
2. The Staleness Crossover Curve
To understand the exact point where standalone models breakdown, we evaluated 134,306 target swims spanning 3,005 swimmers across seven distinct staleness brackets, comparing standalone longitudinal models against pure cross-event imputation:
| Staleness Horizon | Target Event Last Swum | Standalone Longitudinal (t₁) | Standalone Zero-Shot M11 (t₂) | ATLAS Unified Synthesis | Advantage of Unified Model |
|---|---|---|---|---|---|
| Recent | 1–30 days ago | 1.44% MAPE | 1.70% MAPE | 0.46% MAPE | −68% Error vs Long |
| Quarter-Year | 31–90 days ago | 1.93% MAPE | 1.73% MAPE | 1.40% MAPE | −27% Error vs Long |
| Half-Year | 91–180 days ago | 2.38% MAPE | 1.69% MAPE | 1.48% MAPE | −38% Error vs Long |
| Season Gap | 6–12 months ago | 2.37% MAPE | 1.72% MAPE | 1.55% MAPE | −35% Error vs Long |
| 1–2 Years Stale | 366–730 days ago | 2.80% MAPE | 1.76% MAPE | 1.65% MAPE | −41% Error vs Long |
| 2+ Years Stale | >730 days ago | 3.16% MAPE | 1.70% MAPE | 1.62% MAPE | −49% Error vs Long |
These results reveal an unmistakable Staleness Crossover at approximately 60 to 90 days:
- When an event has been raced within the last 60 days, historical target memory dominates cross-event signals.
- Beyond 90 days, an athlete's current form in other events becomes more predictive than their old target swim alone.
- Crucially, Project ATLAS outperforms both standalone approaches in every single bucket. Even when an event has been dormant for over two years, the historical seed swim still carries vital information: it captures the athlete's permanent biomechanical efficiency in that specific stroke and distance that other events cannot directly observe.
3. The ATLAS Architecture
Project ATLAS introduces an integrated multi-tiered pipeline that fuses temporal history and multi-event energetics:
Multi-Timescale Memory Kernels
Target event staleness is decomposed across multiple physiological half-lives (fast acute taper retention, medium macrocycle stability, and long-term career memory), allowing the model to smoothly discount past seed times.
Historical Event Affinity Extraction
ATLAS isolates an athlete's unique stroke efficiency by measuring their historic relative performance: $$\text{Affinity}_E(t_1) = \text{Score}_E(t_1) - \text{Score}_{\text{repertoire}}(t_1)$$ This captures whether a swimmer naturally over- or under-performs in that event relative to their multi-event baseline.
Observed Engine Progression Transfer
Rather than guessing growth from age, ATLAS calculates the exact engine gain across all other active events: $$\Delta\text{Engine} = \text{Score}_{\text{repertoire}}(t_2) - \text{Score}_{\text{repertoire}}(t_1)$$ It then applies hydrodynamic power-law scaling to translate verified engine expansion into expected target-event velocity.
Dynamic Relative-Freshness Gating
A continuous gating ratio balances target memory against current cross-event form: $$R_{\text{fresh}} = \frac{\Delta t_{\text{other}}}{\Delta t_{\text{target}}}$$ When $R_{\text{fresh}} \ll 1$ (other events are fresh, target is stale), the model shifts weight to the progression bridge. When $R_{\text{fresh}} \approx 1$ (all swims are equally aged), it coordinates both perspectives.
Regime-Separated Specialist Routing
To prevent tree feature starvation, ATLAS routes predictions through dedicated branches: a Known Event Progression Ensemble when target history exists, and a Debut-Calibrated Specialist when an event is truly unswum.
4. Empirical Validation & Benchmark Scorecards
ATLAS was validated across the entire 18-event canonical repertoire under strict athlete-level group holdouts ($N = 134,306$ swims across 3,005 unique competitors). The holdout sets were partitioned with zero swimmer leakage and evaluated across multiple random seeds (Seed 42 with $n = 27,666$ and Seed 7 with $n = 26,567$).
| Evaluation Scope | Test Swims | Standalone Longitudinal | Zero-Shot M11 | Project ATLAS | Within ±2% Target |
|---|---|---|---|---|---|
| All Known Events (In-Meet) | 17,598 | 2.15% MAPE | 1.72% MAPE | 1.28% MAPE | 78.1% of swims |
| All Known Events (Pre-Meet) | 17,457 | 2.14% MAPE | 2.17% MAPE | 1.59% MAPE | 70.7% of swims |
| Stale Known Events (>180d) | 5,775 | 2.60% MAPE | 1.73% MAPE | 1.59% MAPE | 71.6% of swims |
| Long-Stale Events (>365d) | 2,466 | 2.90% MAPE | 1.75% MAPE | 1.64% MAPE | 70.2% of swims |
| Stale (>365d) + Fresh Other (≤45d) | 217 | 3.10% MAPE | 1.69% MAPE | 1.56% MAPE | 72.4% of swims |
| Truly Unseen Debuts (0 prior) | 10,068 | — | 2.46% MAPE | 2.45% MAPE | 55.7% of swims |
Deep-Dive: Stale Distance & Specialty Events
The power of the ATLAS progression bridge is most striking in demanding distance and individual medley events, where athletes frequently experience multi-month racing gaps. Below are holdout results on events dormant for over 180 days:
| Dormant Event (>180d Gap) | Holdout Swims | Standalone Longitudinal | Project ATLAS | Error Reduction | RMSE Time Shaved |
|---|---|---|---|---|---|
| 800m Freestyle | 102 | 2.03% MAPE | 1.11% MAPE | −45% error | −6.21 seconds (15.7s → 9.5s) |
| 1500m Freestyle | 49 | 1.86% MAPE | 1.24% MAPE | −33% error | −8.04 seconds (24.6s → 16.6s) |
| 400m Individual Medley | 128 | 1.95% MAPE | 1.10% MAPE | −44% error | −3.65 seconds (8.6s → 4.9s) |
| 400m Freestyle | 280 | 2.37% MAPE | 1.38% MAPE | −42% error | −4.33 seconds (9.8s → 5.4s) |
| 200m Breaststroke | 301 | 2.59% MAPE | 1.42% MAPE | −45% error | −2.71 seconds (6.1s → 3.4s) |
| 200m Butterfly | 153 | 2.31% MAPE | 1.54% MAPE | −33% error | −1.70 seconds (5.2s → 3.5s) |
5. "Sleeper Event" Breakthrough Discovery
One of the most consequential capabilities enabled by Project ATLAS is automated Sleeper Event Discovery. When a swimmer hasn't raced an event in over 12 months, our empirical data shows that 85.1% of return swims result in a personal best, with a median improvement of +4.06%.
However, coaches and swimmers rarely know which dormant event holds the largest untapped drop. We tested ATLAS's ability to identify major breakthroughs ($\ge 4\%$ drop below old PB) across 2,466 dormant return swims:
| Breakthrough Metric (>365d Dormant Swims) | Standalone Longitudinal | Zero-Shot M11 | Project ATLAS |
|---|---|---|---|
| Correlation (r) with Actual PB Drop % | 0.844 | 0.937 | 0.944 |
| MAE on Predicted PB Improvement | 2.90% | 1.69% | 1.58% |
| Precision Flagging Major Breakthrough (≥4% Drop) | 78.1% | 88.8% | 89.5% |
| Recall Flagging Major Breakthrough (≥4% Drop) | 87.8% | 90.2% | 90.5% |
| F₁ Score on Sleeper Breakthroughs | 82.6% | 89.5% | 90.0% |
With an $F_1$ score of 90.0% and an improvement correlation of $r = 0.944$, ATLAS gives coaches an objective tool to spotlight neglected events where an athlete is mathematically primed for a championship-qualifying breakthrough.
6. First-Time Debuts & Truly Unseen Events
When an athlete has never contested an event in their career ($0\text{-Unseen}$, representing 37.4% of all competition entries), the task changes fundamentally.
Our research revealed a Structural First-Time Debut Tax: athletes racing an event for the very first time swim approximately 1.0% slower relative to their cross-event athletic level than experienced swimmers of identical age and aerobic capacity. This gap stems from pacing unfamiliarity, turn timing, and race-day tactical execution.
ATLAS addresses this through three targeted mechanisms:
- Debut Offset Calibration: Explicitly modeling debut status removes the systematic optimism bias present in un-calibrated models.
- Career Stale Repertoire Recovery: If a swimmer has not raced the target stroke in the past 365 days, ATLAS searches their career archive to recover distant non-target swims in that stroke (e.g. using a 50m Breast from two years ago to inform a 100m Breast debut).
- Biological Growth & Recency Physics: Incorporating longitudinal training recency and adolescent growth velocity parameters prevents static snapshots from under-predicting developing youth swimmers.
7. Practical Applications for Coaches & Athletes
Project ATLAS is deployed directly into MyAthlete's cloud inference architecture, powering features across our platform:
- Intelligent Gala Entry Planning: Coaches can evaluate an athlete's projected times across all 18 events simultaneously, identifying optimal event lineups and qualification targets.
- Automated Dormant Event Alerts: The platform flags high-confidence "Sleeper Events" where an athlete's current training form indicates an imminent major PB breakthrough.
- Deep Integration with Sesh Science: ATLAS provides continuous, accurate aerobic anchors (such as 400m Free velocity for Critical Swim Speed) even when distance events have not been raced recently, ensuring training prescriptions reflect true current fitness.
Experience Project ATLAS in MyAthlete
Explore how MyAthlete synthesises longitudinal competition history, real-time training analytics, and cross-event intelligence to give coaches and swimmers unparalleled racing insights.
Explore Swimmer & Coach Platform →