A Venice tide forecast one day ahead is typically off by about 5 centimetres, and two days ahead by about 7. In a full year of out-of-sample tests at Venice, the best model missed the observed level by 5.22 centimetres of mean absolute error at 12 hours, 5.40 at 24 hours, 6.74 at 36 hours and 6.92 at 48 hours (Duttilo & Lisi, 2026, p.23). That is accurate enough to plan a morning around, and not accurate enough to treat a predicted 95 centimetres as a promise.
The reason the number is that good is that nobody forecasts Venice water levels from the tide table alone. Every operational forecast splits the observed level into an astronomical part, which the moon and sun make predictable years ahead, and a weather-driven residual that has to be forecast like weather. At Venice the astronomical part on its own carries 12.06 centimetres of mean absolute error a day ahead; adding a model of the weather residual cuts that to 5.40, a reduction of 55.2 per cent (Duttilo & Lisi, 2026, p.23).
This page explains how the forecast is built, how fast it decays as you look further ahead, what the sirocco and the bora are worth in centimetres, and why the city deliberately errs on the high side when it decides whether to raise MOSE. For the live numbers, see the daily bulletin page.
How accurate is the acqua alta forecast in Venice?
The clearest published test comes from a study that forecast Venice sea level hour by hour across a full year. The models were trained on 2021 and the first half of 2022, tuned on the rest of 2022, then run on 2023 as an untouched test year, producing 8,760 forecasts at every horizon from 1 hour to 120 hours (Duttilo & Lisi, 2026, p.20). That design matters: the figures below are not the model's score on data it had already seen.
For the best-performing specification at Venice, the mean absolute error was 5.22 centimetres at 12 hours, 5.40 at 24 hours, 6.74 at 36 hours and 6.92 at 48 hours. The root mean square errors, which punish large misses harder, ran 7.19, 7.40, 8.98 and 9.20 centimetres over the same horizons (Duttilo & Lisi, 2026, p.23).
In practical terms, a bulletin that says 90 centimetres for tomorrow morning is usually within about 5 centimetres of what the gauge will read, and occasionally out by 15 or more. Venice came second of the six tide gauges in that study for absolute accuracy, behind Nagasaki and ahead of Trieste, Vardo, Nikiski and Saint-Malo (Duttilo & Lisi, 2026, p.23). Part of that ranking simply reflects the small scale of the Venetian tide: there is less water movement to get wrong.
Why the tide table alone is not enough
Venetian water levels are modelled as two things added together. The astronomical component is the rise and fall the moon and sun produce, reconstructed from a short list of harmonic constituents with fixed periods: the principal lunar semi-diurnal at 12.42 hours, the principal solar semi-diurnal at 12.00, the lunisolar diurnal at 23.93, and five others (Duttilo & Lisi, 2026, p.11). The non-astronomical component is everything the weather does on top, and it has to be forecast separately (Duttilo & Lisi, 2026, p.10).
At Venice the astronomical half is modest. The dominant lunar constituent has an amplitude of roughly 23 centimetres (Duttilo & Lisi, 2026, p.12), against a 90 centimetre level at which the lowest parts of the city start to take water. Across 2021 to 2023 the hourly series averaged 32.64 centimetres with a standard deviation of 28.20 and a maximum of 106 (Duttilo & Lisi, 2026, p.8). Those three years contained no exceptional event.
So the common assumption that Venice floods because it has a big tide is backwards. The tide is one of the smallest in that six-station comparison, well under the 684 centimetre mean at macro-tidal Saint-Malo. The weather is what does the damage, which is why a printed tide table cannot tell you whether St Mark's Square will be wet. Modelling the weather residual halves the one-day error, from 12.06 centimetres to 5.40 (Duttilo & Lisi, 2026, p.23).
How far ahead can acqua alta be predicted?
Accuracy decays with the horizon, and it decays fastest early. Across every model tested, forecast errors rose with the prediction horizon, with the sharpest increase inside the first two days (Duttilo & Lisi, 2026, p.23). Between 24 and 48 hours the Venice mean absolute error grows from 5.40 to 6.92 centimetres, so a two-day-old number is worth meaningfully less than a same-morning one.
There is a point where the astronomical tide table stops helping at all. At the 48-hour horizon the harmonic model performed markedly worse than the hybrid models at thresholds above roughly 65 to 70 centimetres, which is precisely the range where the decision to flood-proof a shop or raise a barrier gets made (Duttilo & Lisi, 2026, p.28). Two days out, in the levels that matter, the tide table alone offers limited support.
Operationally the target is about three days, because several days are needed for the timely management of the MOSE system (Goglio et al., 2026, p.4). Tested against a real event, that works. Re-forecasts of the 12 November 2019 acqua alta initialised at 00:00 UTC on 10 November predicted it three days in advance, with errors of 18 and 25 centimetres against the observed 175 centimetre peak, roughly 10 and 14 per cent (Goglio et al., 2026, p.5). The event was called; the exact height was not.
What sirocco and bora are worth in centimetres
The weather residual is mostly wind and pressure, and both have been priced. At wind speeds above 10 metres per second, south-easterly winds (the sirocco) raise the non-astronomical component by about 2.5 centimetres on average, and north-easterly winds (the bora) by about 1.4 centimetres. Westerly and north-westerly winds push the other way, flushing water out of the lagoon and pulling the level down (Duttilo & Lisi, 2026, p.14).
Those are average partial effects across three years, not the peak contribution during a storm, so they describe the ordinary tilt rather than a flood. Frequency matters too. At Venice the wind blew from the north-east in 23.9 per cent of hourly observations and from the south-east in 12.6 per cent across 2021 to 2023 (Duttilo & Lisi, 2026, p.9). The bora is the more common wind; the sirocco, aligned with the long axis of the Adriatic, is the more dangerous one.
Air pressure works through the inverse barometer effect: low pressure lets the sea stand higher, high pressure presses it down. The modelled effect crosses zero at roughly 1015 to 1020 hectopascals (Duttilo & Lisi, 2026, p.14). If a weather app shows a reading well below 1015 while a southerly wind is running up the Adriatic, the two are pushing the same way, and that is the combination the bulletin is watching for.
Why the forecast deliberately errs high
A forecast error at Venice is not symmetrical in what it costs. Overestimating means an unnecessary barrier closure; underestimating means a flooded city. Each MOSE activation is estimated at around 200,000 euros in energy and personnel costs, and the barrier was raised 28 times across 2022 and 2023, roughly 3.2 million euros in total (Duttilo & Lisi, 2026, p.26). Against that, the November 2019 event has been put at 93 million euros of losses plus 82 million in emergency response and structural damage (Duttilo & Lisi, 2026, p.26).
The damage curve bends sharply. One published estimate puts losses at 0.56 million euros for a level of 81 to 90 centimetres, 6.97 million at 91 to 100, 23.03 million at 101 to 110 and 69.08 million at 111 to 120 (Duttilo & Lisi, 2026, p.28). Twenty centimetres near the threshold changes the bill by an order of magnitude, which is why a cautious call is cheap and a brave one is not.
The caution is measurable. Predicted maximum water levels issued by the municipal alert system before each closure have often substantially exceeded the levels actually observed in the city, a systematic overestimation that points to precautionary management rather than to changing weather. That gap has been narrowing over time, even though the operational approach remains conservative (Michielotto et al., 2026, p.848).
Where the closure threshold sits, and why near 90 centimetres
MOSE is a set of 78 hollow flap gates across the three lagoon inlets, raised when levels are predicted to exceed a safety threshold of 1.10 metres above the Punta della Salute datum at Venice (Michielotto et al., 2026, p.848). At that level, flooding reaches roughly 12 per cent of the city area, an extent the city authorities consider acceptable and handle with footbridges and local measures (Michielotto et al., 2026, p.848).
The interesting question is whether that trigger is set in the right place, and the forecast error is what decides it. Running a year of real forecasts through the damage curve and the closure cost gives a loss-minimising threshold of roughly 90 to 95 centimetres for a 24-hour decision window and 80 to 85 centimetres for a 48-hour one, lower at the longer horizon because the forecast is less certain there (Duttilo & Lisi, 2026, pp.28, 29). Both sit close to the 90 centimetre high-water reference level the city already uses (Duttilo & Lisi, 2026, p.29).
That is a useful independent check. The arithmetic of damages, closure costs and forecast error, worked out with no reference to Venetian practice, lands on about the number Venice chose. It also explains why closures begin well below 110 centimetres: the gates have to be committed hours ahead, on a prediction that carries 5 to 7 centimetres of error, for a level the water has not reached yet.
