A personal best is memorable because it compresses a long process into one number. A full running record tells a different story: what was normal, what was exceptional, when performance changed and whether the apparent breakthrough became a new baseline. Across 156 parkruns from August 2018 to August 2026, summarised in the distribution above, that distinction is clearer than it was in the original 118-run analysis.
The central argument remains the same. The 23:31 personal best in June 2021 was not a lucky isolated result. It arrived inside a concentrated sequence of improving runs. The additional 38 results strengthen the other half of the argument: a breakthrough can be genuine without becoming permanent capacity. Form, age-grade performance, season and uncertainty continue to move where the performance level sits. Without continued work and focus on improvement, the former peak does not become the floor; it remains a reminder of what you are capable of.
The whole distribution gives the personal best meaning
The histogram opening this piece places every run in the story. The mean is 28:09 and the median is 28:07, so the repeatable centre of the record is remarkably clear. Most performances sit around that central band, while the 23:31 personal best remains separated in the fast tail.
This is more honest than beginning with the record time alone. A personal best describes the outer edge of demonstrated performance; the distribution describes what was repeatable. Progress should be read through both.
Mid-2021 was a sequence, not a single Saturday
Figure 2. The full chronological record shows the 2021 breakthrough as a concentrated sequence and the latest block as a slower phase.
Every recorded time was plotted chronologically and smoothed with a robust LOWESS curve. That makes a sustained phase visible without pretending that the week-to-week noise of individual results has disappeared.
The trajectory falls through early and mid-2021, while the individual points cluster in the low-24-minute range and culminate at 23:31. Several improving performances create evidence of form before the fastest result arrives. The new results do not disturb that interpretation: none of the additional 38 runs is below 25 minutes, their fastest is 26:36 and their mean is 29:35.
The chart cannot name the cause because the registry contains no training load, sleep, injury, weather or course-condition variables. It can still reject one interpretation. The personal best did not appear randomly inside an otherwise ordinary block; it was part of the strongest sustained phase in the record.
Age grade confirms that the phase was genuinely strong
Figure 3. Age grade independently confirms the 2021 peak, while later improvement remains visible without recreating it.
Each recorded age-grade percentage was plotted on the same timeline and smoothed in the same way as the raw times. This checks the 2021 phase against a second measure rather than trusting the clock alone.
Age grade asks a different question from raw time by comparing performance with an age-related benchmark. It does not replace the clock, but it helps distinguish faster running from the changing context in which that time was achieved. The record-high 54.85% occurs on the personal-best date, strengthening the case that the period represented more than a favourable stopwatch result.
The later record is also informative. A 52.45% result in 2025 creates a smaller second peak, but the best age grade in 2026 is 45.50%. Improvement can recur without reproducing the earlier peak exactly, and a slower phase does not make the original breakthrough less real.
Month changes the baseline, but year changes the runner
Figure 4. Average time and sample size by calendar month across August 2018 to August 2026; lower bars indicate faster performance.
Every run was grouped by calendar month regardless of year and averaged. The sample size above each bar makes clear how much evidence supports each monthly mean.
May remains the fastest month in the record at 27:08, while September is the slowest at 28:59, a difference of one minute and 51 seconds. March is also relatively quick at 27:41, while January averages 28:49. The expanded record has narrowed the apparent seasonal gap from the earlier version of the analysis.
The bars should not be read as a weather law. Each month combines different years, courses and versions of the runner. The exceptional 2021 block still pulls late-autumn and early-winter averages down, while slower phases lift other months. Month provides context; the multi-year trajectory explains why the same month does not produce the same performance.
The forecast respects the distance between peak and baseline
Figure 5. A weekly-state ARIMA forecast and 5,000 simulated paths show the likely 26-week range and the exceptional sub-25 tail.
The forecast follows the same basic approach as the original article. The latest known result is carried forward between attended Saturdays to create a continuous weekly performance state. An ARIMA(1,1,1) model is fitted to that sequence, then 5,000 correlated paths are simulated across the next 26 weeks.
The central expectation settles around 29:28. About 6.6% of the complete simulated paths contain at least one sub-25-minute result. That is the relevant probability for the question being asked: not how many forecast weeks fall below 25 minutes, but how often an entire six-month path contains at least one such breakthrough. The widening bands are as important as the centre because individual parkruns remain variable.
This extends the statistical pattern already present in the record. It says that another sub-25 is possible, but the 2021 peak should still be treated as exceptional rather than assumed to be the current baseline.
Progress has more than one clock
Figure 6. The updated evidence board brings the article's five principal views into one argument.
The board shows why a personal best cannot summarise a running history. The full distribution establishes the baseline, the 2021 sequence identifies a genuine performance phase, age grade offers an independent check, monthly patterns expose seasonal context and the forecast separates likely outcomes from exceptional ones. Each view answers a different question about improvement.
Read together, the larger record argues for patient interpretation. Fast months matter, but so do attendance patterns and the level from which they were achieved. Progress is best understood as a shifting range with occasional breakthroughs, not as a straight line that every future run is obliged to follow.
Progress is a range with occasional breakthroughs
The additional data leaves the original conclusion intact and makes it more secure. The distribution defines normal performance. The time series identifies the concentrated breakthrough. Age grade confirms its quality and shows later movement in context. Monthly averages reveal seasonality entangled with form. The forecast quantifies how uncertain a return to the fastest tail remains.
Running progress is not a staircase in which every personal best becomes the next permanent floor. Consistency builds a range, conditions and training occasionally align to produce a breakthrough, and the record afterwards continues to vary. Later runs are not failures because they do not repeat the fastest one.
The coherent theme is patient accumulation. Most Saturdays establish the baseline; a few reveal what the accumulated work can produce. The refreshed model talks me further out of expecting another sub-25 soon: 6.6% is not zero, but it is much closer to an exceptional outcome than a likely one. That is what a personal analysis should do: give hope a realistic scale, then leave the decision about what to do next to the runner.
How it was done and original sources
The supplied workbook contains 156 complete event rows. Its Excel dates were converted to calendar dates and its displayed clock values were normalised to elapsed seconds, producing the clean CSV used by the replacement notebook. The analysis combines a histogram, robust LOWESS trends for time and age grade, calendar-month averages, and an ARIMA(1,1,1) model with 5,000 simulated 26-week paths. Python, pandas, statsmodels, seaborn, matplotlib and Pillow generate the tables, figures and final evidence board.