Humor & Insight from a Collegiate Cyclist
Velo Quips Get in Touch

A statistical model for predicting podium finishes in Monument races

Monument races sit at the very top of professional cycling, and picking who will stand on the podium at Milan-San Remo, the Tour of Flanders, Paris-Roubaix, Liège-Bastogne-Liège, or Il Lombardia is notoriously difficult. The events are decided by a brutal mixture of parcours, weather, form on the day, and team tactics, which is why so many pre-race tipsters get it badly wrong. Predictive analytics has crept into other sports for years, but applying a structured model to the Monuments is still a relatively niche pursuit. This piece walks through how such a model can be assembled, what inputs it relies on, and where it quietly falls apart, with a few nods to Australian cycling along the way.

The logic behind any credible model is straightforward on paper and unglamorous in practice: collect the right data, weight the right variables, and let the numbers speak. For cycling, that means blending historical results with contemporary form, course characteristics, and a few harder-to-quantify factors like rider psychology and team depth. Anyone who has watched a rain-soaked edition of Roubaix knows that raw data alone will never capture the chaos of cobblestones and gutters. Even so, a well-built model can produce probabilities sharper than gut feel, which is the least any serious punter should expect.

Choosing the variables that actually move the needle

A Monument model needs to separate signal from noise, and that starts with the input variables. Some factors look important on paper but barely move the prediction, while quietly obvious ones do most of the heavy lifting. Course profile sits near the top of any sensible list, since the five Monuments reward very different rider types. Milan-San Remo favours the sprinters who can survive the Poggio, while Liège-Bastogne-Liège belongs to the puncheurs who can summit the Côte de la Redoute and the Côte de la Roche-aux-Faucons in the finale.

Form over the preceding 30 to 60 days is the next layer, often measured through recent race results, UCI points accumulation, and ranking movement. Power data has become more accessible through wattage databases and Strava segments, though a model built purely on raw numbers tends to over-weight outliers who have ridden only one good week. Tactical variables are trickier: team strength in the peloton, protected leader status, and the depth of a squad's lead-out train all influence who reaches the business end of a Monument with help in tow.

Weather is the wildcard that derails most forecasts, and it is especially relevant for northern European events like the Tour of Flanders and Paris-Roubaix. Crosswinds in the Flemish Ardennes, snow on the Muur, or greasy paving stones in the Forest of Wallers can rewrite the script entirely. A model that ignores meteorology is fragile by design, so feeding in a forecast range and running the prediction across several scenarios produces a more honest estimate. A small dose of humility about the weather alone can save a model from overconfident predictions.

Assembling the data without losing the plot

Data quality is where most cycling models quietly collapse. Public race databases provide the backbone, covering historical podiums, finishing times, and stage results stretching back several decades, but gaps in rider biographies and equipment choices mean certain columns are sparsely populated. Aggregating that information into a clean training set requires decisions about what to drop and what to impute, and those decisions quietly shape the output far more than the algorithm itself.

Rider classifications matter just as much. A blanket one-day specialist tag flattens distinctions between a Roubaix rouleur and a Liège climber, two riders who would never realistically share a Monument podium. Splitting the field into archetype buckets, sprinter, classics rider, puncheur, all-rounder, and Grand Tour GC, lets the model score each rider against the specific demands of each race. Course archives, including elevation profiles, sector maps for cobbles, and the location of decisive climbs, anchor the typology in physical reality rather than vibes.

For an Australian lens, the dataset can also track riders who came through the domestic scene. Names that surfaced in the Herald Sun Tour, the Tour Down Under around Adelaide and the Barossa, or the Cadel Evans Great Ocean Road Race near Geelong often belong to riders who later contend abroad. Calmer, longer-form efforts like the National Road Series in towns such as Ballarat and Buninyong have launched more Monument candidates than most casual fans realise. Treating those feeder races as credible early indicators, rather than dismissing them as minor events, sharpens the model for emerging talent before they land on the WorldTour radar.

The modelling approach itself

A handful of statistical techniques are well suited to this kind of problem. Logistic regression offers an interpretable baseline, predicting the probability that a rider finishes on the podium from a small set of weighted variables. Random forests and gradient boosting handle interactions between factors more flexibly, capturing the idea that form matters more on cobbled courses than on flat terrain. A Bayesian framing has its own appeal, since prior beliefs about established favourites can be updated as new races roll in week by week.

Weighting historical results is another subtle but important choice. A win from two seasons ago should count for less than a podium from the current campaign, especially given how quickly equipment, training methodologies, and team dynamics evolve. Exponential decay, which progressively discounts older results, is the most common workaround. The model benefits from rolling updates rather than a single annual retraining run, because Monument racing rewards riders in the middle of their peak windows and punishes those who are slightly off the boil.

For readers wanting a starting toolkit, the key variables worth tracking look something like this:

The model should be calibrated to output probabilities rather than certain podium names, which forces users to confront uncertainty. Saying that Mathieu van der Poel has a 41 percent chance of winning Flanders is more useful than declaring he will absolutely win, especially for anyone wagering on the result or building a fantasy team.

Putting the model through its paces

Validation matters as much as the construction phase. A common approach is to backtest the model against previous editions, predicting podiums with only the data that would have been available at the time. That exercise tends to expose which variables were over-weighted and which were missing. It also reveals where rider form around the spring Classics is more predictive of Monuments than Grand Tour form, a useful nuance for anyone building their own version at home.

The 2023 and 2024 seasons provided a useful stress test. Van der Poel's Flanders wins and Tadej Pogačar's breakthrough at Liège both fitted the model's higher-probability bracket, while the unpredictability of Milan-San Remo continued to trip up forecasts. Paris-Roubaix remained the hardest event to model cleanly, partly because crashes and mechanicals still play a bigger role than form. In a tighter validation window, the model would have correctly named a podium contender roughly half the time across the spring calendar, which is comparable to informed expert tipping and well above random guessing.

Limitations remain, and they are worth spelling out:

Reading the model through an Australian lens

Australian cycling fans have every reason to care about Monument modelling even if no home rider has stood atop a podium for a while. Riders emerging through local clubs, weekend bunch rides, and development pathways often carry the same physiological traits as Monument contenders, and tracking them through a domestic lens helps benchmark what podium-level performance actually looks like at every level. Race organisers in regional Victoria, the Adelaide Hills, and the Sunshine Coast in Queensland increasingly run events that feed into that pipeline, often with timing and parcours that mirror the rigours of northern European classics.

Locally, the conversation tends to skew pragmatic. A mate at the coffee shop in Brunswick or Surry Hills will reckon a model is only as good as the assumptions baked into it, and they are not wrong. Australian cycling culture has always prized a fair dinkum assessment, and that scepticism travels well into statistical work, where cold-eyed evaluation of inputs usually beats optimism. The same approach is helpful when applying predictive models to racing, since overconfident forecasts age poorly when the cobbles start rattling teeth.

Plenty of readers will want to see this kind of analysis pushed further, applied to Grand Tour stages, or even reverse-engineered for amateur racing. There is real appetite locally for sussing out whether a midweek criterium in Geelong or a hilltop finish in the Adelaide Hills can be modelled in similar terms. With more granular data flowing from power meters and race timing chips, future versions of the model could incorporate amateur inputs and produce predictions across the whole depth of the sport.

Plenty more analytical pieces live across the site, and the veloquips.com/about page lays out the editorial direction of the blog and the kind of stories it runs from week to week. Have a look around, share your own podium picks in the comments, and bring the model's logic with you next time you watch a Monument unfold. Predictions always land better when they have been argued over with mates first, and the spring Classics are the perfect excuse to keep that tradition alive.

If you want to test the model against your own instinct, line up a Monument watch with friends, jump on a local group ride the morning of the race, and see how the predicted podiums hold up against the chaos of the real thing. Early starts from Melbourne to Perth have made the European spring something of a national ritual among Australian fans, complete with coffees on the go and chatty group chats tracking the action. Get the data, line up the early alarm, and let the cobbles do the talking.