I tracked all my worked hours for over a year. This (will be) an analysis of that.
https://www.averychan.site/Toggl-Analysis/
I want to try to run a classification algorithm on the times to see if it's able to pick up any patterns in schedules.
I eventually want to cross reference with other sources of information. I used google calendar daily. I also have an app Qbserve that automatically tracks and categorizes computer usage.
For personal reference, I also have journal entries of varying detail to reference to see what effects my productivity or to see what interesting patterns there are.
Since we now have the superpowers of AI, I had Claude run on this again and it had interesting outputs:
Here's the surprise: your data shows the exact opposite of what you expected.
You hypothesized negative autocorrelation — "work 12 hours, crash the next day." But your data shows strong positive autocorrelation: r = +0.37, p < 0.000001. High days are followed by high days, and low days are followed by low days. This persists after controlling for day-of-week (β = +0.374, p = 0.00007), so it's not just a weekday/weekend artifact. The day-of-week controls barely changed the coefficient at all (0.371 → 0.374), meaning your weekly schedule isn't driving this — it's a real momentum/persistence effect.
For every extra hour you work today beyond your day-of-week average, you tend to work 0.37 more hours tomorrow (not fewer). Your work comes in streaks — productive periods cluster together, and low periods cluster together. Look at the time series plot: you can see extended runs of high activity (especially spring 2020) and extended lulls. Even two days ago predicts today (β = 0.16, p = 0.006 in the two-lag model), reinforcing the streak interpretation.
A few things could explain why you feel a crash after a 12-hour day even though the statistics don't support it: The outlier days you remember most vividly (the 12+ hour days) might genuinely be followed by drops — but those are rare events, and the overall pattern across 514 days is dominated by momentum. You can see this in the scatter plot: the huge mass of points in the 0–8 hour range all trend upward. A few dramatic crash episodes are memorable but don't define the statistical trend. It's also possible the rebound effect is real but operates on a weekly timescale (you push hard for a week, then have a lighter week), which would show up at lag 7 rather than lag 1. The ACF plot shows significant positive autocorrelation at lag 7 too — so even at the weekly level, it's momentum, not compensation.
"Daily work hours exhibited significant positive lag-1 autocorrelation (r = 0.371, p < 0.001, n = 514), indicating that above-average work days tended to be followed by above-average work days. This persistence effect held after controlling for day-of-week fixed effects (β = 0.374, SE = 0.093, p < 0.001, HC1 robust SEs), and was present in both 2020 (r = 0.339) and 2021 (r = 0.276) individually. A two-lag model found that hours from two days prior also positively predicted today's hours (β = 0.164, p = 0.006), consistent with multi-day momentum in work patterns." One note: you have some entries over 24 hours in a single day (max looks like ~35h), likely from timers that started before midnight. You might want to clean those if you re-run this — though they're rare enough that they probably aren't driving the result.
