Dew, Data and the Second Innings: Why Cricket's Chase Models Stumble in the Stadium

Dew, Data and the Second Innings: Why Cricket's Chase Models Stumble in the Stadium Last Australian summer I was watching a T20 match at the Sydney Showground. The side batting second needed 62 runs from the last five overs, with five wickets in hand and two set batters at the crease. At the end of the 14th over, my own chase-probability model gave them a 78 percent chance of winning. They lost by nine runs. After the match, at midnight, I opened the log file. The model had assumed the pitch would behave identically in both innings. That was the fundamental error. That night dew had fallen, the ball was wet, and the grip was slipping out of the spinners' hands. The number did not lie; the context behind the number had changed. 'The model said one thing; the empty stadium said another' — I first wrote that line for football, but in cricket it is truer still, because in cricket the pitch is itself a moving variable. At the centre of this piece sits a methodological question: why do T20 chase models keep stumbling in the same place, and how can we test that. Any T20 chase-probability model rests on four variables: required run rate, wickets in hand, balls remaining, and the historical strike rates of the batters still to come. On a large league dataset, those four build a baseline that works quite well. But a baseline is an unconditional average, and the reality of a cricket ground is never unconditional. What are the conditions? Pitch type — slow, turning, or batting-friendly; outfield speed; wind direction; humidity; dew probability; daylight versus floodlights; and, the one we keep forgetting, the age and abrasion of the ball in the second innings. Each of these conditions creates a deviation from the average. My tracking method stays simple. I keep a separate ball-by-ball log for every match, then cross-check it against humidity and temperature data from the venue's weather station. Time, dew probability, pitch report — I attach a 'condition tag' to every innings from those three. Then I ask how accurate the model stays within a single condition tag. This method keeps teaching me that a model's error is rarely in its maths; it sits in the classification of its inputs. When I built my first expected-goals model in a Sydney bedroom in 2026, I learned one thing: 'I do not trust a number I cannot trace to a touch.' I now apply that rule to cricket. Every number must be traced back to its source touch. How much dew, how much pace off the surface, how much the ball has been scuffed — without measuring those, a chase model is just a pretty hallucination. Now to the data. Over the last three Australian summers I have tracked roughly 310 chasing innings across the BBL and international T20s. Alongside watching the games, I kept the ball-by-ball data of the final eight overs of each innings separately. The first pattern that surfaced was this: compared with the first innings, run rate and strike rate in the second innings both rise systematically, especially in the closing eight overs. In my tracking, the average strike rate in the last eight overs of the second innings was about 11 percent higher than the same overs of the first innings. Whichever way the match went, that gap stayed near-constant. Two possible causes sit behind it. First, the chasing side knows exactly what it needs — a target is an anchor, and with an anchor the risk calculation becomes clear. Second, dew. The physical effect of dew is the biggest player of all. When the ball gets wet, seamers lose grip, the ball skids on for spinners, and batters can play through the line instead of waiting for the ball to sit up. In my measured dew index, where intensity is high — night humidity above 85 percent and little grass on the pitch — second-innings spin strike rates were about 14 percent higher than in the first innings. This is the first gap in chase models. Most popular models treat pitch behaviour as fixed across both innings. In reality the pitch changes over time — it is a dynamic system, not a static one. Once dew arrives, the entire equation of the game shifts, yet the model carries no trigger for that shift. The second gap is the misreading of wickets in hand. Models generally assume more wickets in hand means more safety. But late in a second innings, wickets in hand do not equal safety, because strike-rate variance in the lower order is far greater. My data shows the largest gap between model forecast and actual result — roughly 18 percentage points — appears when the chasing side has five or more wickets in hand but a required rate above 10. The third gap is game state. In a first innings, once a side reaches 150-160, it often de-risks in the last two overs, because a par score is now protected. In a second innings the reverse holds. As the required rate climbs, the chasing side takes on more risk in the final two overs. So the last two overs of each innings are two different games. Train a model on both innings in the same bucket and it is measuring apples and oranges together. One data point worth adding here. Second-innings powerplay scoring was only 3-4 percent higher than the first innings — dew's effect on the new ball is limited, because dew has not fully arrived. But after the 16th over the gap jumps above 15 percent. Dew's effect is non-linear; it intensifies at the back end. A model that treats dew as a constant misses that curve. There is a subtler pattern too: dew's effect grows with the age of the ball. With a new ball the humidity difference between pitch and ball is small, so the dew advantage is limited. As the ball ages it gets more scuffed, and in the second innings that older ball gets wetter on a damp outfield. So dew peaks between the 17th and 20th overs. That explains why chasing sides so often overperform in the last four overs, and why live prices move fastest in those same overs. The bowling side needs scrutiny as well. When dew falls, seamers' yorkers and slower balls lose effectiveness, because grip drops on a wet ball. In my data, seamers' economy in the last four overs of the second innings was about 1.3 runs higher than the same overs of the first innings, with wickets per ball down roughly 22 percent. Dew helps the batter and punishes the bowler at the same time. No fixed-pitch model can capture that asymmetry. An example. In that Sydney match, when the spinner came on in the 16th over, the ball skidded off the wet grass after release. In the data, the spinner's revolutions that over were below normal, yet the batter's contact point was further in front. The result — 17 runs off the over. To the model it was a bad ball; to the ground it was a dew ball. That is the difference. And through all of this drifts a misconception: batting second is always easier. On a dewy night that is true, but on a dry, turning pitch batting second is hard — especially if the spinners have already worked the surface in the first innings. For a Rashid Khan-type spinner, a dry pitch makes the second innings mean extra turn, and for a Glenn Maxwell-type finisher a wet pitch means extra advantage. The easy-or-hard question is not set by innings order; it is set by ground conditions and bowling matchups. Format difference matters too. In ODIs dew behaves differently, because the ball runs for 50 overs and in a day-night match dew usually arrives in the last 10-12 overs. In T20s that window covers roughly a third of the innings. Same dew, two different effects. A model that transplants T20 dew patterns into ODIs is getting it wrong. Now to the market, because I work as a betting analyst. Live chase models do not stay inside broadcast graphics; they set prices in the market. Modern cricket markets now include blockchain-based prediction markets, where on-chain smart contracts trade match outcomes. Those markets carry a weakness: they price off broadcast probabilities, and broadcast probabilities often come from that same dew-blind model. The result is one price before dew falls, another after, even though the actual situation has not changed — only our knowledge of it has. On a blockchain market this late adjustment often creates a trading edge, if you can read ground conditions faster than the market. My rule here is clean: 'A transfer rumor is a prior; the medical is the posterior.' In cricket terms — the broadcast probability is the prior, the ground condition is the posterior. Now a warning I apply to my own analysis. Dew and rising second-innings run rates are correlated, but correlation is not causation. 'Small samples are loud; large samples are honest.' The problem is that we reliably measure dew in very few matches. Most broadcasts carry no dew sensor, no record of ball weight and moisture. So we often use the 'dew' label as a proxy — floodlights, humidity, time. A proxy is an estimate, and a model trained on estimates can easily get trapped in its own story. One more thing I test against myself. I often want to over-weight dew, because it fits my frame. But at some venues the bigger cause of second-innings advantage is not dew, it is the first-innings side's excessive caution. Decisions, not only nature. In the model-versus-ground fight, the ground is not always on nature's side; often it is on the side of human decisions. So I keep a threshold: I call a chasing pattern a signal only when it holds across at least 50 matches and two or more venues in the same direction. Below that, it is noise to me, not analysis. There is a practical lesson for readers here. Whenever you see spinners suddenly becoming expensive in a live second innings, ask — did the pitch change, did dew fall, or is the bowling matchup wrong? Each has a different fix. The first is ground truth, the second is weather truth, the third is a selection error. Only the third can be fixed by changing the person responsible; the other two demand a change of strategy. So what will I track next season? One thing — how aggressively chasing sides attack in the powerplay. If a side holds back in the powerplay hoping for dew, it is trusting the model's average, not the ground's condition. And one question to leave you with: when your model shows a 78 percent win probability, do you know which pitch, which humidity, which over that number belongs to? If you do not, then it is not a probability — it is an estimate, neatly arranged. The ground will catch it anyway.