The Statistical Significance of Getting Strong

mildly nsfw · tap to reveal

I have been training, on and off, since about 2013. For most of that decade the consistency was patchy: a handful of workouts in JEFIT (remember that?), some loose change in the Android notes app, and long stretches of nothing at all. Having gone from student to full-time desk-bound drone, I drifted from the university gym to the blue fluorescence of Jetts, learning (badly, at first) to juggle work and a life.

Then, in March 2022, nudged along by one too many GymTok edits (we’re all gonna make it brah) and freshly released as the last of the lockdowns lifted, I started logging every set in Strong, and I have not really stopped since. The gyms have changed again (from the red overtone of Virgin Active to the muffled crimson of Fitness First); the constant has been the log. That log, four years and close to ten thousand sets, is the dataset for this essay.

Progressive overload (see Currier et al., 2026) is the notion that you grow by asking more of the body each week, and it is rather trivial to state. Living it is another matter, because progress is oddly hard to see. The body changes slowly and the exercises constantly. Life intervenes: holidays, work, love, and a bad knee. And fatigue actively masks fitness, so that the days you are strongest often feel the worst. Training without records is like driving around with astigmatism: you can see the haze of lights, but you cannot read any of the road signs. Keep logs for four years and the real signal still doesn’t surrender itself easily - digging it out is what we’re doing in this essay.

One caveat before we start. A training log records what I lifted, never what I looked like, so this is a story about strength, the load on the bar, which is only ever a proxy for size. The two travel together, but they are not the same thing, and the barbell is the one my data actually saw. When I ask whether I am getting stronger, that is the question the log can honestly answer.

I have written it on two tracks. For the gym reader, each section asks what the data says about training and holds it against the modern literature. For the machine-learning reader, it explains the method doing the work, with a little maths kept in expandable asides so you can go as deep as you like. Read either track, or both.

Before any of the modelling, though, it helps to simply see all of it at once, and, let us be honest, who does not love a good contribution graph.

A Gym-Goer’s Odyssey

To turn that calendar into a trajectory, I need a single number for how strong I was in any given week. A bench press and a preacher curl are not measured on the same ruler, and I almost never test a true one-rep max (while fun, it is generally counter-hypertrophic). So for every working set I estimate a 1RM from the weight and the reps (much more on how badly the standard formulas disagree, in Chapter 03), then standardise each exercise against its own history and average across the main lifts.

The result is a strength index. It is unitless: zero is my four-year average, and plus or minus one is a standard deviation above or below it.

Unitless makes it tidy but hard to feel, so to keep it concrete, here is roughly what one point (one standard deviation) is worth in real weight on three of my lifts:

So when the 2024 crunch later costs me 1.18 points, that is about 16 kg off my chest press, or 21 kg off my deadlift.

Plotted raw, that index is jittery, because any single session is the true strength value plus noise: sleep, caffeine, whether the bar felt right on the day. To recover the signal I pass it through a Kalman filter (see Kalman, 1960), a method that treats my real strength as a drifting hidden state and each session as a noisy view of it. It is the same idea that keeps a GPS on course, pointed here at a more modest target.

The change-points that carve the timeline into four chapters (marked by the vertical shading) were not placed by hand. An algorithm (see Killick et al., 2012) scans the series for a few moments of behavioural change, and it recovered the epochs of my life I lived.

Interestingly, we recover a journey. There is a long toil (ponoi) through 2022 and 2023, as raw effort slowly compounds. Then a plateau (stasis), where the line stops rising. Then a fall. And finally the climb back toward excellence (arete), the thing the whole endeavour is quietly reaching for.

The events reflect my life. I started dating my girlfriend in February 2024. That winter, work got genuinely busy, a four-month crunch that, in an earlier era, would have ended my training entirely. This time it did not: the line sags but ultimately holds.

Can we do better than eyeballing that dip? There is a causal inference tool called the interrupted time series (see Bernal et al., 2017). The idea is simple. Take the trend the line was on before the crunch and let it coast onward, a dotted “what would have happened if work had stayed quiet”. Then fit a version of the line that is allowed exactly one kink at the event date, September 2024, and measure how far the real line steps away from that projection. If the step is large next to the ordinary week-to-week jitter, then the crunch, and not chance, is the likely cause.

Figure 3 makes the method visible. In the left panel, the dashed line is the pre-crunch trajectory projected onward, the strength I would plausibly have had if work had stayed quiet; the solid line is what actually happened; and the red gap between them is the loss. It is a large one: an immediate step down of 1.18 standard deviations in the strength index. The fitness rationale is obvious: fewer sessions, thinner sleep, and no mental capacity leftover from building agents all day.

And then, in 2026, the comeback is not merely a recovery. The right panel anchors the same procedure at February 2026: this time the informative quantity is the slope, and the real line pulls away above its flat counterfactual, bending upward by 0.05 standard deviations per week, the steepest sustained strength climb in the last four years.

From 2013 to 2021, a stretch like the 2024 crunch would have ended my training for a year; this time it did not. As Confucius stated: it does not matter how slowly you go, so long as you do not stop.

All the tools here required my personal input. The next chapter builds a model to infer those stages itself.

Epochs, According to the Machine Spirit

Suppose I never told a model about my story, and simply let it watch two numbers each week: how much I trained, and whether my strength was rising or falling. Could it work out, on its own, that training moves through a few recurring stages, and say which one I was in at any given moment? This is exactly the job of a hidden Markov model (see Rabiner, 1989), and it is the natural sequel to the change-points we used before, because it replaces their hard borders with a probabilistic view of each stage every week.

The premise is that behind the messy weekly numbers sit a small number of hidden states, and here I ask for three, which turn out to correspond neatly to building, maintaining, and detraining. The model discovers three clusters (akin to unsupervised clustering) given nothing but two weekly numbers, and I named them afterwards, from their character. Building weeks carry high volume and rising strength; detraining weeks, the opposite; maintaining sits in between.

Two things stand out. The first is that the model, given nothing but two weekly numbers, rediscovers a similar story to Chapter 01: the green building band swells through 2022 and 2023 (the ponoi), collapses into red during the 2024 crunch, and surges back in 2026 (the arete). Its transitions sit close to the dashed change-points, but not on them. Where the change-point algorithm was forced to name a single week, the HMM lets the belief cross over across a month or two, which is how the seasons actually change, by degrees rather than at a stroke.

The second is insightful: the model keeps surfacing periods I had forgotten about. That block of red in early 2023 is when I started a new job, and training evaporated for a while. And 2025, which I remembered as a sedated year, reads on the ribbon as a near-unbroken stretch of detraining, considerably worse than my reflection.

Rosy retrospection is a highly documented cognitive bias. Memory, as Freud liked to insist, is less a recording than a reconstruction, quietly edited in our own favour. Fortunate for me, I’m held to account by the power of numbers and unlimited cloud storage.

Of the 226 weeks logged, the model’s best guess is detraining for 121 of them, more than half, against 76 building and 29 maintaining. When I do train, I exonerate myself and train hard; the trouble is that hard is not the same as often, and the ribbon is really a map of frequency. If I had to bet on the single thing throttling my progress, it would not be intensity but rather the red that is volume. We will put the hypothesis, that volume is what matters most, to a proper test in Chapter 05.

So far, what we have learned is that the model can tell us when I was building or detraining. But the next question is: what did that building amount to in definitive, real kilograms?

Am I Strong?

To ask whether all that building actually worked, I have to put a number on the bar, in kilograms this time, not standard deviations. Two issues get in the way:

1. I almost never test a true one-rep max, so the number itself has to be estimated.
2. More interestingly, even a perfect number is meaningless in isolation. “Strong” is always strong compared to someone, and the answer swings wildly depending on whom you pick.

1RM estimate

Given a set of, say, 60 kg for 8 reps, a one-rep-max formula guesses the single heaviest lift you could have managed. There are many such formulas, most dating to the 1980s and 1990s, and no two of them agree; each was fitted to its own author’s athletes or a modest sample, so each bakes in a slightly different guess about the same curve. At one rep they must all return the weight on the bar, but as the reps climb they fan apart, because extrapolating from a set of twelve up to a hypothetical single is genuinely uncertain. I computed seven of them; Figure 5 shows how far they drift.

To pick one, I checked each formula’s estimate against the rare days I actually ground out a heavy single or double: O’Conner and Lombardi came closest (about 10.6 and 10.9 kg off), while the popular Epley and the exponential Wathan overestimated the most. Those two winners are the estimators behind every kilogram that follows. The full bake-off, formulas and all, is in Appendix C.

Relative strength

Now that I have my one-rep maxes, how strong am I? For years I religiously checked strengthlevel.com, which takes a lift, quietly routes it through this same kind of formula, and hands back a flattering verdict: novice, intermediate, advanced, and a percentile. Vainly, I felt a sense of accomplishment as my lifts crept up the tiers. It was answering one specific question, “strong compared to the general population of people who log lifts”, and to that question my 2.24 times bodyweight deadlift is a respectable intermediate.

Tougher crowds exist, and against them the verdict flips. OpenPowerlifting (OpenPowerlifting, 2026) publishes the meet results of people who actually compete, and who therefore presumably train specifically for strength. Filtering to the 66,000 male raw lifters within five kilograms of my bodyweight and dropping my lifts onto their distribution (Figure 6) gives a very different picture. The same deadlift that strengthlevel calls intermediate lands at the 24th percentile among competitors, and my bench at the 40th. Blue is everyone I out-lift; grey, the majority who out-lift me.

Somewhere between the flattering mirror and the merciless one sits a third, more considered measure. In How Strong Should You Be? (Noob to Freak), Jeff (Nippard, 2024) builds a scale from a decade of coaching and competition data, sorting lifters from noob through intermediate and advanced up to the genetic freak. On his ladder my 2.24 times bodyweight deadlift lands right on the intermediate-to-advanced rung: better than OpenPowerlifting’s crowd made me feel, worse than strengthlevel let me believe, and probably the most honest of the three.

The video’s sharpest insight, though, is not in the video at all but in its top comment, which compresses the entire problem of reference classes into a great mindset:

Nevertheless, the resonance is the same in every percentile: “how strong am I” turns out to be less a question about me than about whom I choose to stand beside. And unless you are chasing a competition total, the only person you should be standing beside is the version of yourself who trained last. You just have to beat him a little, and often enough.

At this point, I would love to do a strength analysis based on the exercise: a chest press, a row, a curl. The issue is that endless variations exist: dumbbell benches and Smith benches and inclines, machine rows and cable rows. Before any exercise’s strength level can be trusted, those scattered variants have to be put into a common currency.

The Ship of Theseus

The Greeks liked to ask whether a ship whose planks are replaced one by one, until not a single original board remains, is still the same ship. Assume my bench press now resembles a ship.

I have pressed it with a barbell, with dumbbells, on a Smith machine, on an incline, on a chest-press machine, and on whatever the new gym happened to have bolted to the floor, and wasn’t taken. Not one of those is the lift I started with, and yet, in every way that matters to me, it is all still “chest press”. The plates change, the angle changes, the hardware changes. The muscle underneath does not much care (for the most part).

This is a pleasant philosophy, and a headache for measurement. Every time I swap variants, the naive record resets. A dumbbell that tops out at 32 kg per hand and a Smith bench that moves 80 kg are not remotely the same number, so a chart of any single exercise is a string of false stalls: a lift climbs, gets abandoned for its cousin, and the cousin appears to start over from nothing. Look at exercises, and I plateau constantly. Instead, if we look at the general movements, the gains are there.

Stitching the planks together

The trick is that swaps are rarely clean breaks. For a week or two I usually do the old exercise and the new one side by side, and those overlaps are the seam. If, across the weeks I did both, my incline bench sits reliably at about 1.15 times my flat dumbbell bench, then I have a conversion rate between them, an exchange rate in the currency of estimated one-rep maxes. Pin one variant as the anchor, chain those exchange rates across every overlap, and all fifteen presses collapse onto a single scale. The stitched line each week is simply my best set that week, converted onto that one shared scale.

The trap of the absolute axis

There is a catch though with absolute weights. Even benchmarked to a barbell curl, my biceps top out under 40 kg, less than half of what my chest press moves, so on a shared kilogram axis they hug the floor and look like they never budged. This makes plotting difficult.

The fix is to stop reading absolute kilograms and re-base every movement to 100 at its first week, so the y-axis becomes percent of baseline. Now a 19 kg curl and an 86 kg press are asked the same question, how far have you come, and answer on the same scale. Far from flat, my biceps turn out to have ratcheted up the most of anything, about +180 percent at their best, with the honest caveat that a tiny starting number flatters the ratio.

Each solid line in Figure 8 is that best-to-date ratchet: records only rise, so the line only climbs, and every step is a movement PR. Re-based this way the chapter’s point finally lands. My biceps, dead flat in kilograms, ratchet up as hard as anything I train.

It is a reassuring picture, though the first draft of it was not. Every climb here rests on a pile of conversions between variants, and on the raw rates a few of those conversions were simply wrong, enough to show my arms drifting backwards in the months I was training them hardest. I am, in fact, stronger than ever. So before trusting the chart, I had to check the rates it stands on; that full audit lives in Appendix A, with the short answer being: where overlaps exist the rates are solid, but 17 of the 61 were never measured directly and had to be filled in from typical population ratios.

Conclusively, variation is, for the most part, fine. Modern work on exercise selection is fairly relaxed about the exact tool, and increasingly interested in where in the range of motion the tension lands: lengthened, stretched-position training holds up well against full-range work (see Wolf et al., 2025), and the technique reviews treat variant choice as a lever for emphasis and joint comfort rather than a make-or-break decision (Androulakis-Korakakis et al., 2024). My drift toward inclines and machines was not self-sabotage; it reflected that slow, fashionable reshuffle.

You can watch the ship get re-planked. Split every kilogram I have ever lifted by the tool that moved it, quarter by quarter, and the mix never holds still. I came up on free weights: barbells and dumbbells were 71 percent of my volume at the very start. By 2024 they had ceded the floor to the machines and cables, which together crossed sixty percent of everything I did. The 2026 comeback dragged some barbell work back in, but the arc is unmistakable, a slow migration from the free-weight rack to the pin-loaded stack. Every one of those swaps is a plank replaced; the movement underneath, and the strength I built on it, carried on regardless.

Volume: How Much Is Enough?

I have counted many sets. My training has been a slow tour of other people’s programs, each a different bet on how much work is the right amount. I came up on StrongLifts 5x5, all low reps and heavy frequency; drifted into the bodybuilding gospel of 4x12 and its wall of pump; and I have mostly settled into a hard, honest 3x6–8, three sets taken close to the edge and not much else, though I will note that 2x8 is now all the rage. The through-line, if I am honest, is that I have been slowly buying less volume and paying for it with more effort. This chapter asks whether that was wise, and whether the dose even does what everyone says it does.

The framework the set-counters swear by comes from Renaissance Periodization (see Israetel et al., 2021), which gives every muscle three weekly landmarks: the MEV, the minimum effective volume you need to grow at all; the MAV, the adaptive band where most of the productive work happens; and the MRV, the maximum you can recover from before the returns go negative. They are not laws of physics. They are heuristics, distilled from the hypertrophy dose-response literature and a lot of coaching, individual and fuzzy at the edges. The modern meta-regressions broadly back their shape: more sets keep helping, with diminishing returns rather than a cliff (see Pelland et al., 2025), which sharpens the older picture (Schoenfeld et al., 2017). The one thing to hold onto is the unit. These are weekly numbers, per muscle.

Flip through the muscles and a confession assembles itself. Upper body sits just over the minimum; legs barely register. Hamstrings are trained in fifteen percent of weeks. Leg day is mostly what I skip (haha), and it is the clearest thing I need to fix.

Does the dose actually work, though?

The whole framework assumes volume causes growth. My logs can show that, but beware the old statistician’s adage: correlation is not causation. Weeks when I do more sets are also weeks when I am heavier, already stronger, better rested, and in a better mood, and any of those could be doing the work while volume takes the credit. Simply correlating sets with next-month gains would hand the prize to whichever lurking cause shouted loudest.

A rigorous way to check for causation is to use a causal graph (see Pearl, 2009): weekly sets point at strength gain, but thirteen other variables point at both, so they have to be held still. I control for prior strength, modelled fitness and fatigue, intensity habits, workload ratio, mood, and flu, to assess whether volume is really a driver. But volume is really two levers: more sets within a session, and more sessions within a week. To separate them I put both in a joint model, each controlling for the other.

Figure 11 shows both overlaid. The sets slope (blue) goes negative; the days slope (aqua) stays steep at +0.16 points of the strength index per extra training day. Sets and days are highly correlated, so it takes this joint model to tell them apart: it is not how much I do when I show up, but how often I show up.

Two caveats keep this honest:

1. The effect is calculated from my own noisy life, not a controlled trial, so it is strong suggestion, not proof.
2. While thirteen confounders shut most of the obvious doors, a lurking one I forgot could still be whispering.

If I wanted to prove it I would run the experiment: randomly assign each coming block to high or low frequency, hold everything else fixed, and read off the difference. That would be a real trial. I humour these ideas, but they stand in the way of gains.

There is a blunter way to see the same thing. Skip the causal machinery and just plot the dose of volume against the strength payoff: bin every week by how much I did, and read off the strength change that followed. Figure 12 settles the chapter’s question in two curves. More training days keep paying off, climbing until about four or five a week. But sets piled into a single session peak near fifteen (about five exercises for me) and then bend negative, the point past which more work in one sitting quietly costs me. Frequency is a dose I can keep raising; per-session volume is one I cannot.

My median is three sessions a week. I have run Push-Pull-Legs before, and I am running it again now, but my history is mostly a loose Upper-Lower-Upper where the second Lower rarely happens. The joint model says the fix is clear: get beyond three days. The extra session buys more than extra sets ever did, and by the time I am deep into a long workout I am tiring out anyway, so keeping per-session volume relatively lean and spreading the work across more days is the honest next move.

The Cult of the Last Rep

In 2022 my feed was one long sermon on intensity. Every set went to the death, because a rep in reserve was a rep half wasted. I chased drop sets for the burn, bought the lengthened partials, the whole liturgy. The gospel was that it was not how much you did but how hard, and I took it to heart. Four years of records let me test the faith.

The session view

The first thing the data reveals is how devout I was: I take at least one set to failure in seventy-nine percent of sessions. That near-universal habit wrecks the obvious test, because my rare no-failure days are deloads and walking-wounded sessions that were going to PR less anyway. So the honest question is not whether I train to failure but what share of each session I do, and how that proportion tracks two things: a record that day, and what is left in the tank afterwards.

A little failure helps: pushing from none to a tenth of my sets lifts my record chance from 48 to 66 percent. But grind more than a third to failure and the rate falls back to 54 percent, and the next-session strength change sinks from +0.07 to −0.09 points (roughly a kilo on the bench either way). Drop sets are kinder: the record rate climbs steadily and the next-session line stays flat, no fatigue penalty. The extra volume at reduced load appears to be genuinely free. Some of that failure bump is a mirage (I redline harder on days I already feel strong), but the saturation and the fatigue cost are real. The literature agrees: failure adds little over stopping a rep or two short, and it costs recovery (see Grgic et al., 2022), with newer work tracing the same curve for proximity to failure (Robinson et al., 2024) and for hypertrophy (Refalo et al., 2023).

The month view

A single session is only one stopwatch. Fatigue is short-lived, so it is worth zooming out. Take the trailing month’s failure proportion, lagged by one week (a lag just shifts the window back, so the current week cannot predict itself), and follow strength one to four weeks later.

Lighter-failure stretches surge to +0.09 points (about a kilo on the bench) by week four; months of grinding a fifth or more of every set to failure bank almost nothing (+0.05 points, half that). The payoff comes from the submaximal work, not the redlining.

Not every muscle agrees

“Failure” does not mean the same thing to every muscle. Split each group’s weeks into its heavier- and lighter-failure halves and follow it a month out: back and shoulders shrug the redlining off, even edging ahead on heavier weeks, while arms and quads wilt. Triceps are the extreme, 0.17 points worse (a couple of kilos on a pressing lift) when I grind a big share to failure.

The pattern is suggestive but physiological: big compound movers bank the effort, small isolation muscles drown in fatigue. “Train to failure” was never one instruction but a dial to set differently for a heavy row than for a cable pushdown.

The prescription

The last rep is not holy. A little failure is useful, it’s how I grind through to a PR, but a lot of it is a cruel central nervous system tax with little return. Drop sets are the smarter tool: they sneak in extra volume at reduced load without the fatigue penalty, and the record rate keeps climbing with the dose. So the smart thing to do? A rep or two left in the tank on most sets, use drop sets more frequently, and save the death-grind for the rare day. Am I going to listen? No.

The Roads Not Taken

Two things happened to me that no amount of good programming could out-argue. In 2024, a brutal, months-long stretch of overtime and burnout, when there was simply nothing left after a day of building agents. And in late 2025, a meniscus, after which I quietly stopped training legs and kept hammering everything in a Johnny Bravo fashion. To ask what each cost me is to ask a counterfactual question: not what happened, but what would have, on the road I did not take.

Pricing a road not taken means building it, a plausible version of me who was never interrupted, and measuring the distance to the one who was. The catch is that my two shocks have different shapes, and different shapes need different scaffolding.

The 2024 crunch was pervasive. It hit everything at once, so there is no healthy corner of my training to borrow as a control, every lift sagged together. When the whole system moves, you have to become your own control: fit the trend from before the crunch and let it run on untouched, an interrupted time series (see Bernal et al., 2017), and read the toll off the gap between that projection and reality. We already watched this happen in Chapter 01.

The meniscus was the opposite shape: contained. Only my legs went down; my chest and back and arms carried on regardless. That is the textbook setting for a synthetic control (see Abadie & Gardeazabal, 2003), which builds the missing counterfactual out of the parts that were spared. So I fit my leg pressing against my healthy upper body over the years I have records for both, then let whatever relationship there is predict what it should have done through the injury and after. The dashed line below is that prediction. The solid line is the hard-hitting truth.

And the chart refuses to tell the story I wanted it to. If the injury were the villain, the two lines would track together and split at the red marker. Spoiler: they don’t. They begin peeling apart in early 2024, well over a year before the knee gave out. By the eve of the injury my leg press already sat some 38 kilograms adrift of where my upper body put it; the meniscus, for all the blame I heaped on it, added under ten more. Chapter 05 already confessed as much: I had been quietly abandoning my legs for a very long time, and the knee merely handed me a respectable alibi to finish the job. This was a reality check I didn’t expect.

Both shocks were real, but only one was external. The crunch was something that happened to me; the leg deficit was something I did to myself, one skipped session at a time, long before the knee gave me permission to stop entirely. The counterfactual machinery can price an interruption, but it cannot manufacture the data you never bothered to collect. The lesson is less about causal inference and more about honesty: if I had trained legs consistently, the injury would have had something to interrupt.

You Are Fitter Than You Feel

There is a familiar cruelty to the day after a brutal session. The bar that flew up on Monday feels welded to the floor on Tuesday, and the obvious reading is that I have gone backwards. I have not. The truth is kinder: I am fitter than I was on Monday and, at the same time, more tired, and what my hands report on the bar is not my fitness but the difference between the two. This is why progress is so hard to feel from inside a training block. The thing you are building and the thing that is masking it are the same act of training, running on two different clocks.

Every hard set does two things at once. It builds a slow, durable fitness that accumulates and fades over weeks, and it drops a fast fatigue that piles up in a day and clears in a few. Strength on any given morning is the slow thing minus the fast thing. Rest for a week and fatigue evaporates while fitness barely moves, so you walk in feeling exactly as strong as you have quietly become. Lifters call this a taper; the accidental version is a “deload PR”.

Fit that model to four years of my sessions and you can finally watch the two clocks run. Figure 17 draws them on the same strength scale I have used throughout: the blue line is what my strength would be on fitness alone, the ink line is the net form the model actually predicts, and the red band between them is the fatigue tax, thickening for a few days after every heavy stretch and then melting away.

One thing in that picture is worth being honest about: the red band is invisible. When I fit the model, it settles on two dials: one for how much lasting fitness each session lays down, and one for how much short-term fatigue it drops. Mine come out lopsided. Every hard session deposits more than three times as much durable strength as it costs me in tiredness, and because that tiredness drains away within a week while the fitness lingers for six, almost none of it is left to show up as a fatigue tax on the bar. At my fittest the fitness term is worth roughly thirty kilograms of bench and fifty of deadlift over a fully detrained version of me. The fatigue term, even on my most buried day, subtracts barely two kilograms of bench, four of deadlift. The model captures only about a third of the story, and what it does capture is almost all durable fitness rather than a fatigue mask waiting to be lifted.

Either I do not accumulate fatigue the way the textbook athlete does, or I am permanently fatigued and have simply adapted to it. I have always had high endurance and a tolerance for pain and physical discomfort, and I wonder if this is itself a neurological adaptation, the kind of thing that four years of grinding through sessions has quietly trained into me. Deloading, for me, according to this model, does nothing.

The Banister model was built for athletes who peak and taper, who accumulate real fatigue debt and then cash it in before competition. That is not me. My training is steady, my fatigue is negligible, and the model’s explanatory power is modest. But what it does reveal is reassuring: nearly everything I have built over four years is durable. The slow memory holds. The fitness is real, and it is not hiding behind a wall of tiredness waiting to be unmasked. It is simply there, compounding quietly, session after session.

The Reckoning

Every chapter so far has pulled back one veil. We filtered out the noise, compared 1RM formulas, looked into volume, failure, fitness and fatigue. What if we looked at it top-down, without prejudice, and let a model decide for itself what earns a record?

What actually predicts the day I set a personal record?

So I handed a machine everything I could know going into a session, the whole recent history behind it: rolling averages of volume, mood, flu levels, and temperature, each computed over the last week and the last month. Then I asked it a simple question at widening horizons, from the next session out to six weeks: are more records coming than usual? The model is a machine-learning method called XGBoost (Chen & Guestrin, 2016); to read its mind I use SHAP, which tells me, for every single session, which way each feature pushed.

The honest answer: barely, and only for tomorrow

Before I trust a single feature, I have to ask whether any of this is predictable at all, and I have to ask it honestly. That takes three disciplines. First, a target that is not secretly a calendar. Asking “did I beat my all-time median PR count” sounds fair, but because my PR rate has drifted over four years, it really just asks “is it 2026?”. So instead I ask “did I set more records than my own recent baseline, the last two months”. Second, a validation that respects time: train only on the past, test only on the future, and never let the model peek at a session that shares an era with its training data. And third, a noise floor. I scramble the log so each session keeps its real training history but is handed some other session’s outcome, then run the model again. Whatever it scores on that nonsense is pure luck, and any real result has to beat it.

Scored that way, the verdict is sobering. I grade each model with a measure called AUC, where 0.5 is a coin flip and 1.0 a crystal ball. The next session carries a faint but real signal, 0.55, just clear of the 0.49 luck line. Three days out it is already fading, and from a week onward the model just bobs around inside the luck band, a month ahead included. The grey band in Figure 18 is that luck floor, and every horizon past the first lives inside it.

I sat with that for a while, because I wanted it to say more. Four years of diligent logging, every set and every mood, and the machine’s honest answer is that tomorrow is mostly a mystery and next month is a closed book. Progress, it turns out, is no easier to forecast than it was to feel. But there is that one stubborn flicker at the very next session, so let me follow it.

What faintly moves the next session

So there is a whisper, one session out. What is it made of? Start with the levers I actually pull (Figure 19). Honestly, not much: hard, high-volume grinding tilts the odds a little, failure and drop sets both nudging toward a record, alongside the durable fitness I have banked. But the whole controllable picture clears the noise floor only barely, 0.55 against a 0.49 luck line. The things I decide move the needle, just faintly.

Two of those tilts glance off earlier chapters. Failure pointing toward a PR sits awkwardly against chapter six’s verdict that it is mostly a tax, but the two ask different questions: that chapter measured the causal dose, where a little helps and a lot costs, while this model only sees that I fail sets most in my hard, engaged stretches, the same ones that throw off records. It reads failure as a marker of the grind, not a cause. Fitness, by contrast, points exactly where chapter eight said it would: the durable base is what carries a record, and my fatigue is so slight that a fresh day has nothing to add.

None of it is loud. Now switch on the things I can’t control and the model climbs to 0.59, but look at what carries it. Not the weather, momentum. The single biggest tell in the whole model is how many records I have set in the last month, and a hot streak is not a lever, it is an outcome. Records cluster, so the surest forecast of my next one is just that I have been setting them lately. Temperature muscles in behind it, and I really do set fewer records in the heat, as the next section shows, but that tall bar flatters it: the honest accuracy barely moved, because a gentle seasonal tilt adds little you can bet on a session out. What little I can foresee about tomorrow’s record leans mostly on things I do not decide.

Because SHAP is additive, I can watch the full model work on a single day. Figure 20 takes a real record from March 2024 and pulls its call apart. It opens at an average session’s 59 percent and climbs to 75 percent, but look at how. The named factors nearly cancel: a strong last session, a week of hard failed sets, and a decent mood push up; a late hour and a cold streak push back. What actually does the lifting is everything else, the long tail of small features that, summed, outweighs every headline bar combined. Even on a confident day the model is not pointing at a reason; it is adding up a hundred faint rumours, which is what a weak signal looks like from the inside.

What I can’t: flu and heat

Those last two, flu and heat, deserve a closer look, pulled as they are from real Australian influenza surveillance (WHO FluNet, 2026) and real Melbourne weather (Open-Meteo, 2023). They leave their fingerprints in opposite places: flu on the door, heat on the bar.

Flu works on attendance. When it is circulating I show up a little less, but cross the threshold into the gym and it vanishes: my record rate in high-flu weeks is 62.1 percent, and in low-flu weeks, 62.1 percent, identical to the decimal. Flu decides, faintly, whether the session happens, never whether it succeeds.

Heat is the mirror image. It barely changes whether I show up, but it tracks how I lift. I set records far more often in cool weeks than in the heat, 70.6 percent against 53.5 percent, and the gap holds in every year of the log, not just my beginner surge. Whether it is real physiology or just that I grind harder when I am not melting, Melbourne’s summers are quietly my leanest season for records. That seasonal fingerprint is why the model reached for temperature a moment ago, though the tilt is gentle enough that betting on it a session ahead barely helps.

What I feel: the diary

I had assumed that how I felt, captured in three hundred and seventy-one scribbled notes and scored by a sentiment model (RoBERTa (Barbieri et al., 2020)), would be pure noise. It is not, and it cuts two ways at once. Over months, brighter sentiment tracks a stronger me: my notes’ mood correlates +0.15 with the strength index, the contented entries landing in my strongest stretches. Yet inside the PR model the sign flips. A sour, frustrated recent month, the very thing that drags that sentiment down, is what faintly precedes a record, because frustration is what a hard grind feels like from the inside. Feeling good is where the work has got me; feeling frustrated is often what the next record is made of. Both pulls are faint, and the model leans on the second only lightly.

The emotions themselves are almost embarrassingly transparent. The sentiment model, told nothing about lifting, works out that I am at my sourest about lateral raises and leg extensions, and my fondest about heavy rows and incline pressing. It reads my mood better than I would admit, and my mood, it turns out, carries a faint rumour of the next record.

When: the Monday effect

One last lever, and unlike flu or fatigue it is entirely mine to schedule: when I train. My records are not spread evenly. Monday is my strongest day by a clear margin, 1.41 PRs per session against 0.99 on a tired Thursday, part real recovery cashed in from the weekend, part the fact that I open the week with the lifts I care most about. And because this is a raw count, not a model, it survives everything above. If a day reliably buys a better shot at a record, the attempts belong there.

So the reckoning is not the one I went looking for. I could not build a model that forecasts my records more than a single session ahead, and the one honest signal, that a record tends to follow a hot streak, is a whisper of information. The outside world, too, only reaches me in a soft tone: the flu gently turning me away, the hot months grinding me down. My own diary, even, is a poor predictor. What predicts a PR? Almost nothing, in advance. But the earlier chapters were never really about prediction; they were about the odds. Frequency beats intensity, a rep left in the tank beats training to failure, the legs I keep skipping are the clearest gain I own, and even here Monday is quietly my best day. None of it tells me when a record will land. All of it makes one likelier. So you show up more often than you feel like, leave a rep in the tank, train the legs, do the volume, and go home to quietly progress. You only find it afterwards, and that is precisely why we write this all down.

So, Am I Stronger?

The question the whole essay opened with was a simple one, and after twenty-five figures it finally has a plain answer. Yes. The strength index sits at its all-time high. The comeback is, by the interrupted time series of Chapter 01, the steepest sustained climb anywhere in four years. And stitched across every swapped variant, in Chapter 04, every major movement is now above where it has ever been. The log cannot tell me whether I look any different in the mirror, and I promised early on it never could. But on the one question it can honestly answer, the load on the bar, the verdict is not close.

What did four years actually build? Figure 24 lays it over a body. Colour each muscle first by how much its strength grew since the first week I logged it, and almost everything glows: the presses and pulls doubled, and the legs, on the rare stretches I trained them, grew faster than anything else I own. Then flip the toggle to training volume, the hard sets I actually spent, and the legs go dark. That single switch is the whole confession. My quads and calves posted the biggest gains in the record, flattered though they are by how little they could once do, and they sit at the very bottom of everything I trained.

So I will spare myself the pretense that any of this was a mystery, and just write down what the data has been telling me the whole way through:

• Train legs. It is the clearest fix in the entire dataset. They respond when asked; I simply stop asking, and a meniscus was only ever the alibi (Chapter 07).
• Show up more often, not harder. The extra session buys more than the extra set ever did, and the sets I cram past fifteen in one sitting quietly cost me. Get past three days a week (Chapter 05).
• Leave a rep in the tank. Failure is a tool for a good day, not a tax to pay on every set; drop sets sneak in the same volume without the bill (Chapter 06).
• Bank the records on Monday morning, fresh off the weekend, when both the calendar and the clock are on my side (Chapter 09).

None of it was hidden. It was only hard to see: buried under swapped exercises and bad memory, masked by a fatigue I turned out barely to carry, blamed on a knee that was mostly innocent. I could not feel the progress from inside the training, and I could not remember it honestly from outside. The only reason I can hand any of it back now, as a verdict instead of a vibe, is that four years ago I started writing it all down.

How Trustworthy Are the Conversions?

A currency is only as trustworthy as the trades that set its rates. My stitching learns a rate between two variants solely from the weeks I did both; where two variants never met in the log, the model has nothing to price from and quietly assumes they are equal, a factor of exactly 1.00. Across these five movements, 17 of the 61 rates are these un-measured defaults. When the two variants really are near-identical, a machine pulldown and a cable pulldown, the assumption costs nothing. When they are nothing alike it is a silent lie: a weighted pull-up, which hauls my whole bodyweight over the bar, gets booked as equal to a lat pulldown, and a barbell skullcrusher as equal to a cable pushdown.

To see how far the money has drifted from reality, Figure 25 sets each variant’s learned rate beside a real-world reference, drawn from the population standards that sites like strengthlevel and the lifting subreddits converge on. Pick a movement; its anchor is the big compound of the group, pinned at 1.00.

The pattern is not subtle. Where I trained two variants together often, the rate is trustworthy: my dumbbell curl prices at 0.47 of a barbell curl against a population 0.50, my skullcrusher at 0.79 of my pushdown. Where I did not, it breaks. Seventeen of the sixty-one rates never share a week with the rest of their movement, so the model cannot price them and I backfill from the population instead. Even a measured rate carries asterisks: it is assumed constant, though my skullcrusher climbed from 0.45 to nearly level with my pushdown between 2022 and 2026; an anchor has to be a lift I still do; and a logged number is only load when the weight is the work, which cable stacks and bodyweight dips quietly violate. I did not detrain. I simply never gave the model the overlaps it needed, which is the sharpest lesson in an essay about record-keeping: a log can only stitch what you thought, at the time, to measure side by side.

A few of the measured rates are interesting precisely because they are mine and not the crowd’s. My preacher curl comes in at 0.77 of my barbell curl where the average lifter sits near 0.98, so the preacher, which strips out every scrap of body english and holds the biceps at a long, disadvantaged length, punishes me harder than most. I would like to blame short biceps on a short frame, and the reverse curl, feeble at 0.69, keeps the theme going. My neutral and mag-grip pulldowns come in a little under a plain pronated one, 0.84 and 0.93, which points at grip and forearm rather than lats. I hypothesise that the currency has quietly become a biomechanics probe: wherever my number drifts from the population’s, it is likely fingering a leverage, a limb, or a joint angle that suits me a little worse than average.

Turning a Log Into Data

Every figure in this essay stands on one unglamorous act: turning a phone app’s export into something a model can read. The raw material is a single CSV that Strong hands you on request, about ten thousand rows, one per set, four years deep. What sits between that file and the first chart is the part nobody photographs, so here it is.

Parsing the log

Each row is a set: a date, an exercise name as I typed it, a weight, a rep count, and sometimes a rest timer or a note. The first job is to read it faithfully. Strong encodes warm-ups, drop sets and failure sets in small tags, and it logs rest timers as their own zero-weight rows, so a naive read double-counts them as lifts. The parser splits genuine working sets from rest timers, repairs the occasional hour-long “duration” left running when I forgot to stop it, and rolls the sets up into sessions. Ten thousand rows collapse to roughly eight thousand working sets across 491 sessions.

The naming problem

The real mess is the exercise names. Over four years and four gyms I have typed the same movement a dozen ways: Bench Press (Dumbbell), DB Bench, Dumbbell Bench Press, plus outright typos. Left alone, each spelling is a separate exercise with its own lonely history, which is fatal for any analysis of progress. Collapsing them is a three-step funnel. A rule-based pass parses the structured part of each name into a (base, equipment, grip) key. A fuzzy-matching pass (rapidfuzz, similarity ≥ 94) sweeps up the near-duplicate spellings. And where spelling alone cannot tell that two names mean the same lift, an embedding pass does: each name is turned into a vector by Gemini Embedding 2, and names whose vectors sit close together are matched. The 148 raw names become 132 canonical exercises, with the meaningful variants (dumbbell versus Smith, incline versus flat) deliberately kept apart, because Chapter 04 needs them.

Attaching muscles

A set only knows its exercise, not the muscles it trains, and every per-muscle claim in the essay (the volume landmarks of Chapter 05, the body map of the coda) needs that link. Each canonical exercise is matched, again by embedding nearest-neighbour, to the open-source free-exercise-db of 873 catalogued movements, which carries the primary and secondary muscles, the force direction, and whether the lift is compound or isolation. Volume is then attributed with a primary muscle counted at full weight and each secondary at half, so a bench press credits chest fully and shoulders and triceps in part.

Keeping it honest

Two habits run through the whole pipeline. The first is that nothing physical is invented: weekly series are all anchored to the same Monday so they align, bodyweight lifts are kept out of the load-based stitching because their logged number is not the work done, and the causal claims are put through placebo and refutation tests whose results are reported whether they pass or not. The second is that the outside world is real, not simulated: the percentile crowds come from OpenPowerlifting, the weather from Open-Meteo, and the flu season from the WHO’s FluNet, each joined onto the days it belongs to. The whole chain, from the raw CSV to every JSON that feeds these charts, regenerates from one command, so every number here is reproducible rather than remembered.

Which Formula to Trust

Almost everything in Chapter 03 and beyond leans on a single move: turning a set of, say, 60 kg for 8 reps into an estimated one-rep max. A one-rep-max formula does exactly that, and there are many of them, most dating to the 1980s and 1990s. No two agree — each was fitted to its own author’s athletes or a modest sample — so each bakes in a slightly different guess about the same curve. They coincide at one rep, where they must all return the weight on the bar, and fan apart as the reps climb, because extrapolating from a set of twelve up to a hypothetical single is genuinely uncertain (that fanning is Figure 5).

With $w$ the weight and $r$ the reps, the seven I computed are

They are just different guesses at the same curve, some linear in reps, some hyperbolic, some exponential, which is why they agree near $r = 1$ and part ways as $r$ grows. In order, these come from Epley, 1985, Brzycki, 1993, Lombardi, 1989, O'Conner et al., 1989, Landers, 1985 (the McGlothin/Landers form), Mayhew et al., 1992, and Wathan, 1994; (LeSuer et al., 1997) is the classic head-to-head validation of several against measured maxes.

To decide which one suited me, I evaluated them empirically. Across the lifts where I did occasionally grind out a heavy single or double, I can compare each formula’s estimate, computed from my higher-rep sets, against that demonstrated near-max. Admittedly, reading a one-rep max off a fatiguing multi-rep set muddles fatigue with true strength, which is why every formula overestimates to some degree; but absent a habit of monthly true-max tests, this is the best signal on offer, and it is the relative ranking that matters.

Ranked by average miss, O’Conner and Lombardi win (about 10.6 and 10.9 kg off), while the popular Epley and the exponential Wathan sit at the back, overestimating by the most. That is why those two are the estimators behind every kilogram in this essay.