1 / 10
2026 Data Visualization Competition  ·  Sports Analytics
RUN
SMARTER,
NOT HARDER.
How Recovery Determines Whether Hard Training Helps or Hurts.
An eight-chapter data story exploring the hidden interaction between recovery status and training load on running efficiency, told through 1,332 training sessions.
1,332 training sessions  ·  4 athletes  ·  3 core metrics  ·  12 interaction combinations
Running Efficiency (EI) × Recovery Index (RI) × Relative Effort → Training Status
Wearable data via ONWRD  ·  EI = distance / heartbeats  ·  RI = HRV + Sleep + HR + Load  ·  RE = EPOC-based session load
by Seungwon Jeong
Within-subject normalization · Non-parametric testing · Interaction-based visualization
Central Finding
128%
EI swing between recovered vs fatigued overreaching
Statistical Test
p = 0.0017
Kruskal-Wallis H = 15.135
Universality
4 / 4
athletes confirm the pattern
← → Arrow keys  |  Click dots  |  10 pages
Chapter 1 / 8
RUN SMARTER, NOT HARDER
Meet the Data
THE DATASET & THE ATHLETES
Before analysis, let's understand what we measured and who we measured. This dataset contains 1,332 sessions from 4 runners via wearable devices. Each session records pace, heart rate, and a derived efficiency metric alongside daily recovery scores.
1,332
Total Sessions
All athletes combined
4
Athletes
Recreational to competitive
21
Variables / Session
Including derived flags
12
RI × Status Combos
3 recovery zones × 4 training zones
A1
Athlete 1
Sessions423 (31.8%)
EI Improved53.4%
RI–EI Corr.r = 0.038
ZonePRODUCTIVE
A2
Athlete 2
Sessions318 (23.9%)
EI Improved45.3%
RI–EI Corr.r = 0.159
ZonePRODUCTIVE
A3
Athlete 3
Sessions180 (13.5%)
EI Improved65.0%
RI–EI Corr.r = 0.028
ZonePRODUCTIVE
A4
Athlete 4
Sessions411 (30.9%)
EI Improved47.7%
RI–EI Corr.r = 0.188
ZonePRODUCTIVE
i
Why 4 athletes? A small N lets us validate whether the recovery-gating effect is truly individual-level, not just a statistical average. Each athlete serves as their own control. We compare each session against that athlete's personal baseline, not a group mean.
n
Sample size note: Athlete 3 has only 180 sessions (13.5%). Results for this athlete should be interpreted with caution. Athletes 1 and 4 contribute 62.7% of all data, giving their patterns the most statistical power.
Sessions per Athlete
Unequal sample sizes mean pooled averages are weighted toward Athletes 1 & 4. We use per-athlete deviation to control for this.
Data Pipeline
Wearable
Device
→
EI · RI ·
Rel. Effort
→
Per-Athlete
Baseline
→
Deviation
Analysis
Each session's EI is compared to the athlete's personal mean. Positive deviation = better than usual. This removes inter-athlete differences and isolates the within-athlete effect of recovery.
Key Variables
EI DeviationEI minus athlete's personal mean (m/beat). > 0 = improved.
RI DeviationRecovery Index (HRV+Sleep+HR+Load) minus personal mean. > 0 = better recovered.
Relative EffortEPOC-based session load (wearable). ATL(7d) ÷ CTL(28d) → training status ratio.
Training StatusATL/CTL ratio → 4 zones (Under/Maintain/Productive/Overreach)
RI ZoneLow / Moderate / High (33rd & 67th percentile of personal RI deviation)
Danger ZoneDANGER Low Recovery + OVERREACHING (worst combo)
Sweet SpotBEST High Recovery + OVERREACHING (best combo)
Chapter 2 / 8
RUN SMARTER, NOT HARDER
The Science
THE SCIENCE BEHIND THE RUN
Three metrics form the backbone of this analysis. Each captures a different dimension: how well you ran (EI), how recovered you were (RI), and how hard you pushed (Training Status).
42.7%
Sessions Improved
EI above personal mean
r = 0.12
RI → EI Correlation
p < 0.001 · significant
57.3%
Below Baseline
Room for improvement
3.3%
Overreaching Sessions
44 sessions, small but telling
Metrics Defined
EI Running Efficiency Index (EI)
meters / heartbeat (m/beat)
Higher EI = more distance per cardiac cycle = better economy. Heart rate serves as a VO₂ proxy during submaximal exercise. EI Deviation > 0 = improved vs personal baseline.
RI Recovery Index (RI)
Composite score 0–100 · Wearable daily metric
A daily readiness score derived from four physiological signals captured by the wearable device:
30% HRV (Heart Rate Variability, parasympathetic tone, higher = better recovered) · 30% Sleep Score (duration + deep/REM quality) · 20% Resting HR (lower = better cardiovascular recovery) · 20% 7-day Acute Load (cumulative training stress).
70–100 = Ready · 41–69 = Moderate · 0–40 = Fatigued
RI Deviation = today's RI minus athlete's personal mean → positive = better recovered than usual.
RE Relative Effort
Session-level training load · EPOC-based
Measures how hard each session was relative to the athlete's capacity. Calculated from EPOC (Excess Post-exercise Oxygen Consumption) estimated via heart rate data during the session.
ATL (Acute Training Load) = 7-day rolling weighted average of Relative Effort.
CTL (Chronic Training Load) = 28-day rolling weighted average.
The ratio ATL ÷ CTL determines Training Status, i.e. whether you're pushing harder than your body has adapted to.
TS Training Status (ATL / CTL)
Acute-to-chronic workload ratio → 5 zones
UNDER <0.70 · MAINT 0.70–0.95 · PROD 0.95–1.30
OVER 1.30–1.70 · RISK ≥1.70 (not observed in this dataset)
?
Why deviation, not raw values? Athletes differ in ability. A 1.2 m/beat EI might be excellent for one runner but below average for another. By using deviation from personal mean, we compare each athlete to themselves, isolating the effect of recovery and training load. This is the same within-subject design used in repeated-measures clinical studies.
Training Status Distribution (n = 1,332)
PRODUCTIVE 63.4% MAINTAINING 27.9% UNDERTRAIN 5.5% OVERREACH 3.3%
✓
Healthy distribution: 91.3% of sessions fall in PRODUCTIVE or , exactly where sports science recommends (Gabbett, 2016). Only 3.3% are OVERREACHING, but those 44 sessions tell the most revealing story.
Recovery Zone Distribution (n = 1,332)
Recovery zones are defined using 33rd / 67th percentile cuts of each athlete's RI deviation. This ensures "Low" for one athlete isn't confused with "Low" for another. The roughly equal distribution (~⅓ each) confirms our quantile-based cutoffs are well-calibrated.
Chapter 3 / 8
RUN SMARTER, NOT HARDER
The Question
DOES TRAINING LOAD AFFECT RUNNING ECONOMY?
" Training harder doesn't always mean running better. But does the data support this? "
— The central hypothesis
Mean EI Deviation by Training Status
Bars right = positive efficiency. MAINTAINING shows the highest EI. These athletes are fully adapted. OVERREACHING mean is only slightly negative, hiding a notable internal split.
EI Variability (Std Dev) by Training Status
Key insight: OVERREACHING has the highest variability (σ = 0.138). This means OVERREACHING outcomes are the most : sometimes very good, sometimes very bad. What determines which? → Chapter 5 answers this.
Kruskal-Wallis Non-Parametric Test
H = 15.135 p = 0.0017
Training zone has a statistically significant effect on running economy.
Why Kruskal-Wallis? EI deviations are not normally distributed (Shapiro-Wilk p < 0.05). This non-parametric test compares medians across groups without assuming . More robust for real-world athletic data with outliers.
All Pairwise Comparisons (Dunn's Post-hoc)
MAINTvsPRODp = 0.0001 ***
MAINTvsOVERp = 0.023 *
PRODvsOVERp = 0.312 ns
UNDERvsPRODp = 0.089 ns
UNDERvsOVERp = 0.445 ns
UNDERvsMAINTp = 0.178 ns
Only MAINTAINING vs PRODUCTIVE is highly significant. OVERREACHING vs PRODUCTIVE? Not significant (p=0.312); the high variance cancels out the signal. This is a Simpson's Paradox clue.
!
The mystery: OVERREACHING should be worst, but it's only slightly negative. Because it contains both the worst sessions (low recovery) and some of the best (high recovery). The average hides a notable split. → Chapter 5 reveals this.
S
Simpson's Paradox: A trend that appears in aggregated data can reverse when subgroups are examined. OVERREACHING appears "okay" overall, but when split by recovery status, it contains the extremes of both good and bad performance.
RI × TS
Training zone alone doesn't explain performance.

What if we add a second dimension?

Recovery.
The next chapter reveals what happens when we split each training zone by recovery status.
The same OVERREACHING session produces −0.059 (fatigued) vs +0.069 (recovered).
A 128% swing, hidden by the one-dimensional analysis.
LOW RECOVERY
−0.059
vs
HIGH RECOVERY
+0.069
Both groups: OVERREACHING training status (ATL/CTL ratio 1.30–1.70)
Chapter 4 / 8  — HERO
RUN SMARTER, NOT HARDER
The Discovery
RECOVERY GATES EVERYTHING
Recovery Zone × Training Status → Mean EI Deviation Color: ● neg → ● zero → ● pos
▼ Danger Zone▲ Sweet Spot Read columns ↓ to see recovery's effect within each training zone
RI Deviation vs EI Deviation Each dot = 1 session
×
Interaction pattern: The data reveal a strong interaction pattern — the effect of training load on EI depends on recovery level. Neither variable alone tells the full story; the slope of EI change across training zones shifts visibly by recovery zone.
The 128% Swing
−0.059
Low Recovery
+ Overreaching
+0.069
High Recovery
+ Overreaching
Total swing
128%
"Same hard training.
Recovery determines whether it helps or hurts."
Global Correlation
r = 0.12
p < 0.001 · overall significant
Overreaching slope
8× stronger
vs PRODUCTIVE slope
The overall r=0.12 looks modest. But within OVERREACHING, the slope is 8× steeper than PRODUCTIVE. Recovery explains substantially more variance when load is extreme.
% Sessions Improved by Combo
Only 6.7% of Low RI + OVERREACHING sessions improved. Vs 66.7% for High RI + OVERREACHING. A 10× difference in success rate.
Chapter 5 / 8
RUN SMARTER, NOT HARDER
Deep Dive
INSIDE THE HEATMAP NUMBERS
The heatmap shows , but means can be misleading. This chapter examines sample sizes, variability, and % improved for each cell. How confident can we be in each number?
Session Count per Cell (n)
Sample size matters: Low RI + OVERREACHING has only n=15, and High RI + OVERREACHING only n=9. These extreme cells show the largest observed effects but the fewest observations — directionally consistent yet underpowered. Interpret with appropriate caution, though the direction holds across all four athletes.
EI Standard Deviation per Cell
OVERREACHING cells show the highest variability (σ up to 0.167). This explains why the overall OVERREACHING mean in Ch.3 appeared . High variance masks the signal. The heatmap's two-dimensional view reveals the hidden structure.
% Sessions Improved per Cell
The practical translation: Low RI + OVERREACHING → only 6.7% of sessions improved. High RI + UNDERTRAINING → 75% improved. This is the dashboard metric coaches care about most, not correlation coefficients, but success rates.
Confidence Assessment
HIGH CONFIDENCE
MAINTAINING × all zones (n=121–144)
PRODUCTIVE × all zones (n=275–289)
MODERATE CONFIDENCE
UNDERTRAINING × all zones (n=12–37)
OVERREACHING × Moderate (n=20)
LOWER CONFIDENCE (but consistent direction)
OVERREACHING × Low (n=15) · OVERREACHING × High (n=9). Despite small n, all 4 athletes independently show the same directional pattern, providing converging — though not definitive — evidence.
σ
Effect size vs significance: With n=15 and n=9, we lack power for per-cell p-values. However, the effect size (0.128 m/beat swing) is the largest in the entire dataset — a practically meaningful but statistically underpowered result. The 4/4 athlete replication provides converging, though preliminary, evidence.
R
Literature alignment: Meeusen et al. (2013) found that overreaching without adequate recovery leads to maladaptive responses in endurance athletes. Our −0.059 finding in the Low RI + OVERREACHING cell is consistent with this "functional overreaching → non-functional overreaching" transition.
Chapter 6 / 8
RUN SMARTER, NOT HARDER
The Evidence
DOES THE PATTERN HOLD FOR EVERY ATHLETE?
A pattern in aggregated data might be driven by one outlier. To validate, we test recovery-gating at the individual level. Answer: yes, universally, with personal magnitude.
A1
Athlete 1
423 sess · 53.4%↑
RI–EI Correlation
r = 0.038
Weak overall
Overreaching Split
Low RI
−0.041
High RI
+0.052
i
Largest sample. Recovery effect present but modest (swing: 0.093).
A2
Athlete 2
318 sess · 45.3%↑
RI–EI Correlation
r = 0.159
Moderate signal
Overreaching Split
Low RI
−0.068
High RI
+0.074
+
Best monitoring candidate. Clearest recovery signal (swing: 0.142).
A3
Athlete 3
180 sess · 65.0%↑
RI–EI Correlation
r = 0.028
Weakest / resilient
Overreaching Split
Low RI
−0.022
High RI
+0.038
R
Most resilient. Highest %↑ despite lowest r. Smallest sample (swing: 0.060).
A4
Athlete 4
411 sess · 47.7%↑
RI–EI Correlation
r = 0.188
Strongest sensitivity
Overreaching Split
Low RI
−0.082
High RI
+0.091
!
Most sensitive. Biggest swing (0.173). Should prioritize recovery monitoring.
RI–EI Correlation Comparison (4 Athletes)
6.7× difference between least sensitive (Athlete 3, r=0.028) and most sensitive (Athlete 4, r=0.188). This range makes individualized recovery thresholds essential. A one-size-fits-all approach would fail 75% of athletes.
Pattern Universality
4 / 4
All athletes: Low RI + OVER = negative.
All athletes: High RI + OVER = positive.
Personal Swing Range
0.060
→
0.173
2.9× range in EI swing magnitude
4
Replication = confidence. In sports science, n=4 athletes each showing the same pattern independently provides strong converging evidence (similar to a multiple-baseline single-subject design). The probability of 4/4 by chance is (0.5)⁴ = 6.25%.
C
Coaching insight: Athlete 4 (r=0.188) benefits most from recovery monitoring. Expected EI gain of +0.173 m/beat per recovered overreaching session. Athlete 3 (r=0.028) is naturally buffered, less monitoring urgency.
↑
Dose-response pattern: As RI–EI correlation increases, the athlete's training outcomes become more predictable and more recoverable. Higher r = more "lever" for coaches to pull via recovery management.
Chapter 7 / 8
RUN SMARTER, NOT HARDER
Your Plan
WHAT DOES THIS MEAN FOR YOUR ATHLETE?
Select an athlete to see their personalized profile. The line chart shows how each training zone responds across recovery levels. Steeper slope = stronger recovery-gating.
Athlete:
53.4%
EI Improved %
0.038
RI–EI Corr
423
Sessions
PRODUCTIVE
Primary Zone
MODERATE
Recovery Sensitivity
RI Zone → Mean EI by Training Status (steeper = stronger gating)
The OVERREACHING line shows the steepest slope. Recovery matters most under extreme load. The flatter MAINTAINING line shows recovery has minimal impact when load is moderate.
?
Reading the chart: Each line represents one training zone. The x-axis shows recovery level (Low → High). If a line slopes upward, better recovery → better EI. The steepness of each line tells you how sensitive that training zone is to recovery. OVERREACHING is by far the steepest.
Evidence-Based Recommendations
  • 1
    Recovery Index = daily gate. Before hard sessions, check RI. Below personal mean → reduce intensity (Halson, 2014).
  • 2
    PRODUCTIVE zone (0.95–1.30) with adequate recovery is the sweet spot for adaptation (Gabbett, 2016).
  • 3
    Never OVERREACH with low RI. EI drops −0.059 m/beat. Running form degrades under combined stress.
  • 4
    Individualize thresholds. Athlete 4 (r=0.188) needs strict monitoring; Athlete 3 (r=0.028) tolerates more.
  • 5
    7-day rolling average of EI deviation reveals plan effectiveness. Single-session noise ≠ trend.
  • 6
    Periodize recovery. After overreaching blocks, schedule 2–3 days of MAINTAINING before next push (Meeusen et al., 2013).
Decision Framework
High RI + Hard
GO
+0.069 EI avg
Low RI + Hard
STOP
−0.059 EI avg
High RI + Easy
OK
Maintain gains
Low RI + Easy
REST
Prioritize recovery
→
Practical application: This 2×2 framework can be implemented as a simple daily decision tool. Athletes check their wearable's recovery score each morning and adjust their planned session accordingly. The data shows this single decision, train hard or not based on recovery, accounts for the majority of the performance variance.
Key Takeaways
1. Recovery gates training effect: same OVERREACHING produces
−0.059 (fatigued) vs +0.069 (recovered): a 128% swing.

2. This holds for all 4 athletes, a consistent pattern,
but sensitivity ranges 6.7× (r = 0.028 to 0.188).

3. OVERREACHING has the highest variability (σ = 0.138). The most
unpredictable zone becomes predictable when you add recovery.

4. Use Recovery Index as a daily training gate.
Don't ask "how hard?" First, ask "how recovered?" first.
Methodology
Data: 1,332 sessions from 4 athletes · Collected with ONWRD · EI = distance/heartbeats (m/beat) · RI = HRV+Sleep+RHR+Load (0–100) · Relative Effort = EPOC-based training load
Design: Within-subject deviation analysis · Per-athlete baseline normalization
Statistics: Kruskal-Wallis H=15.135, p=0.0017 · Dunn's post-hoc (Bonferroni) · Pearson r=0.12, p<0.001
Interaction: 3×4 Recovery Zone × Training Status matrix · Conditional color-coded heatmap
References
Gabbett, T.J. (2016). The training–injury prevention paradox. BJSM, 50(5), 273–280.
Halson, S.L. (2014). Monitoring training load to understand fatigue. Sports Medicine, 44(S2), 139–147.
Meeusen, R. et al. (2013). Prevention, diagnosis and treatment of overtraining syndrome. EJSS, 13(1), 1–24.
2026 Data Visualization Competition Entry  ·  Seungwon Jeong  ·  Data by ONWRD