From Harvest Forecasts to Market Readiness: Can Data Reduce Bangladesh's Post-Harvest Losses?

Bangladesh is not, for the most part, losing its harvest for lack of information. Much of the grain rots because there is no dryer, fuel or buyer waiting when the information arrives, and more data is unlikely to change that until a forecast obliges someone other than the farmer to move.

That is the uncomfortable answer to the question in the headline. Data can cut post-harvest losses, but mainly the share caused by not knowing. This year's evidence suggests that share is smaller than the enthusiasm for agri-tech assumes.

The case rests on two harvests. One is the Boro paddy of the north-eastern haor, flooded this April five days after an accurate warning. The other is the 2025 potato crop, grown into a glut that the planting figures had signalled months earlier. Both suggest the binding constraint now sits downstream of the forecast.

The Loss Nobody Has Properly Counted

Consider the size of the hole first. An estimated 34 per cent of the food this country produces is lost or wasted, a drain valued at about 4 per cent of GDP, and 27 per cent of the land grows food that nobody eats. Those figures come from a World Bank diagnostic presented in Dhaka in September 2025.

The damage is spread unevenly across the plate. The same study put net loss rates at 23 per cent for rice and 29 per cent for mangoes, and FAO estimates presented alongside it ran from 17 to 32 per cent for fruit and vegetables. Among the remedies on the table were digital agriculture applications and better early warning for climate events. Sensible advice, all of it. It also rests on an assumption that deserves a harder look, which is that the warning is the weak link in the chain.

Those proceedings carried an irony of their own. A government food-planning official told the same conference that the country still lacked baseline data on how much it loses, and that the statistics bureau had begun a survey to find out. Data is being prescribed for a loss that has yet to be properly measured.

The technology has still earned a fair hearing. Any primer on digital farm forecasting tools will describe weather data pooled in one place and satellites reading a crop's condition from orbit, so that yield can be estimated before a single stalk is cut. Little of that is hype. Growers who use such tools irrigate and fertilise with less guesswork, and the claim that less of their crop is lost along the way is a reasonable one.

Five Days in April

Then April arrived. On the 23rd the state flood forecasters warned, in a bulletin carried by the national news agency that rivers feeding the haor basin could rise above their danger marks for three days from the 28th. That was five days of notice for seven districts that grow about a fifth of the Boro crop. The forecast held. By the first days of May more than 47,000 hectares of paddy were under water, and the extension department still counted a quarter of the haor crop in the fields.

The national figures barely registered it. FAO's latest country brief estimates the 2026 Boro harvest, which supplies about 55 per cent of annual paddy output, at roughly 31.9 million tonnes, close to the previous year's record. A professional reading that number in Dhaka could reasonably conclude the season went well. In a basin that grows one rice crop a year, a household that lost it has no second harvest to fall back on, and the aggregate cannot show that.

What could a farmer do with those five days? On 24 April only 18 per cent of the crop in those districts had been cut. Most of the rest was still days from ripe, and a scientist at the rice research institute put grain in its last stage of filling four or five days from maturity. A field can be warned, but it can't be hurried. The squeeze also appears to be tightening. A study of April rainfall between 1995 and 2022 points to pre-monsoon rain increasing and shifting from May into April, the month of the harvest.

So the advice went out to cut any field that was 80 per cent ripe, which asks a family to accept a certain loss today against a probable one next week. Many tried. Combine harvesters could not enter waterlogged plots. A machine that burns 100 to 120 litres of diesel a day was being fuelled through a token system, and operators who secured a token sometimes found the dealer had nothing to sell.

Hand labour, the old fallback, was missing too. Growers offered half the crop to anyone who would cut it, or day rates of Tk1,500 against the usual Tk700 to Tk800, and reported no takers. The extension department's regional chief gave a reason. Harvesting has become machine-based, he said, and that has thinned the seasonal flow of workers from other districts. Mechanisation raised the basin's speed in a normal year and seems to have removed its reserve in a wet one. The bulletin had done its job, and almost everything downstream of it was improvised.

Where the Forecast Itself Runs Thin

None of this means the forecasting is finished business. A preprint from a team of researchers, most of them in Sylhet, counts fewer than twelve working gauges across the Sunamganj haor, a basin of some 8,000 square kilometres. The paper has yet to pass peer review, and it notes that three of the nearest upstream gauges were out of service during the 2017 flood. Their account is that official bulletins reach district offices for the main rivers while no haor-specific forecast reaches farming communities directly. If that is right, better instruments and better delivery over the last mile would buy real time.

The same paper carries a caution for forecasters and app builders. Alert errors have unequal costs. A false severe warning can push a family into cutting a crop early for nothing, and an early all-clear can leave it standing in the path of the water.

No Sun in the Forecast

The worse failure in April came after the cutting. Paddy leaves the field wet and has to be dried promptly or it sprouts and rots. Sun drying takes two to five days of cooperative weather, and a flash-flood warning is, among other things, a warning that there will be no sun. The alternative already exists. A bin dryer adapted at Bangladesh Agricultural University takes half a tonne of paddy from roughly 25 per cent moisture to under 9 in seven and a half hours. It runs on electricity for the blower and briquettes or gas for heat.

The evidence for it is local and has been in print since 2019. Trials published by its developers measured drying losses of under half a per cent in the machine, against roughly 2.4 to 4 per cent for paddy sun-dried by farmers in Tangail, Mymensingh, Netrokona and Jessore. The authors reckoned that gap, applied nationally, at about 1.4 million tonnes of paddy a year. That is an engineer's extrapolation and should be read as one.

By the extension department's own account, the haor had nothing like that capacity when it mattered. In Khaliajuri, half of what one farmer managed to harvest could not be dried and began to sprout. He had planted nearly ten acres and expected to keep the grain from one and a half at most. Paddy that fetched Tk800 to Tk900 a maund straight from the field in earlier years went for Tk500 to Tk600 when it sold at all. A senior extension official in Sylhet put the position plainly. "We have to depend on nature," he told this newspaper, in a week when the flood centre was still issuing river bulletins three days ahead.

That price gap was set by traders with no instruments at all. Watch one at work and you see moisture judged as strictly as any sensor would judge it, and priced within minutes. Information travels fast in a rural market. The person holding the wet sack simply has nowhere else to take it. State procurement opened on 3 May, and economists were by then urging the government to buy wet paddy, which its depots do not ordinarily accept.

Potatoes and the Price of Last Year's Signal

Market readiness, the second half of the headline, is mostly plumbing. A crop is ready for market when there is somewhere dry or cold to hold it and a buyer who has to quote a price for it. Price apps, on the evidence so far, come a distant third.

In 2025 growers planted potatoes on 5.24 lakh hectares against an official target of 4.67 lakh, so the overshoot was sitting in the acreage figures before the first tuber was lifted. Output reached about 1.3 crore tonnes against demand of some 90 lakh tonnes and cold storage for 45 lakh. By October, potatoes that had cost one Bogura grower Tk28 a kilogram to grow and store were selling for Tk9 or Tk10 at the cold storage gate, and most of one trader's stock was fit only for cattle.

Even the rescue seems to have made matters worse. In late August the agriculture ministry announced a floor price of Tk22 and a plan to buy 50,000 tonnes. By mid-October the price had not been enforced and nothing had been bought, and growers who held their stock in expectation watched the market fall further. An announcement is a data point too, and this one had nothing behind it.

Potatoes are also where the case against data weakens. A planting forecast published early enough might have kept some of that acreage in another crop, and the ministry said last year that it was working with the statistics bureau on satellite imaging and AI to align production with demand. That's data doing what only data can do. Such a forecast earns its keep because the number reaches the grower before his decision is locked in, while he can still choose differently.

Who Still Has a Choice in Late April

Apply that test to the haor and the gap is plain. A five-day flood warning reaches a farmer whose choices were fixed months earlier, when he transplanted, and in every budget cycle that did not put a dryer in his union. The parties who still have choices in late April sit elsewhere. They are the offices that allocate diesel and the agencies that could buy wet paddy at a posted price. They are also whoever decides where harvesters and dryers are stationed before the season opens. For them the forecast is closer to a news item than an instruction, because little in the system appears to require them to act on it.

Which is why the enthusiasm for putting more data in farmers' hands, however well meant, looks aimed at the wrong hands. The farmer tends to act on every signal he gets, including the bad ones. He cuts green paddy when told to. He plants potatoes because last year's price told him to. A country serious about loss would wire its forecasts to obligations, so that a river bulletin of a given severity released diesel and moved harvesters into the basin without anyone convening a meeting.

The country has done something like this before, with this very forecaster. In July 2020 a UN pilot on the Jamuna tied money to two triggers, the second of them the flood centre's own prediction that the river would cross a set threshold at Bahadurabad. When that prediction came, agencies began distributing cash and supplies under a plan agreed in advance, and reached about 200,000 people. The design had three parts, which were a forecast, money committed in advance and released on triggers, and an action plan settled beforehand.

The approach also appears to work. An Oxford working paper on cash sent to Bangladeshi households five days before a forecast flood peak found better welfare during the flood and three months on, lower asset losses, and gains from acting even one day earlier. The hill districts offer a smaller and more recent case, where landslide warnings now arrive with door-to-door volunteers and, for the most vulnerable families, Tk5,000 toward the cost of moving. Humanitarian cash is a simpler thing to trigger than a fleet of harvesters. The principle that a forecast should release something is the same.

A Working List for the People with Options

District and upazila officials hold the cheapest fix, and it should come first. Write the trigger down before the season. Name the gauge and the forecast level, and state what is released when the flood centre predicts that level will be crossed. Diesel heads the list, because April showed that a harvester without fuel is idle for the one week that matters, and a pre-authorised allocation costs little in a dry year. Harvesters borrowed from neighbouring districts come second. A roster of hand-cutting crews belongs on the same page, since machines stall in standing water and migrant labour no longer arrives unasked.

Drying is the business case inside this story, and it belongs to millers, equipment dealers and the banks that finance them. At half a tonne a batch, the university's bin dryer is a village-scale tool, so a union needs several or a mill needs something larger, and that is a lending decision more than an engineering one. Drying sold as a hired service, the way harvesters are already hired out, would let a farmer buy an afternoon of machine time instead of a machine. Buyers can help by posting a moisture discount schedule in place of flat refusal, so that wet paddy has a known price.

Anyone building agricultural software might start with the addressee. An alert is worth sending when the person receiving it still has a move to make, so a flood warning belongs with the diesel allocator and the mill owner at least as much as with the grower. It should also say how ripe the crop is, since ripeness decides what five days can achieve. Delivery deserves equal thought. In the Chattogram landslide programme only 38 per cent of warnings arrived by mobile phone, and the rest travelled by loudspeaker, religious institutions and volunteers. Then count what happens after the cut. Haor farmers complained at the start of May that official tallies recorded submerged fields while ignoring cut paddy rotting in the yards, which is the very loss this debate is about.

Cold storage has its own version of the same step, which is to publish capacity next to acreage while planting is still under way. A grower deciding how many bighas to put under potatoes should be able to see how many tonnes the stores in his district can take. The 2025 figures existed, and there is little sign that anyone set them side by side.

Before the Next Bulletin

All of this would cost money in the years when the water doesn't come. A dryer bought for a flood year sits idle in a dry one, and an auditor will ask why. Officials would be spending against a probability, which is the one habit a forecast can't make comfortable. There, arguably, lies the real limit of data. A forecast describes the risk five days out and leaves open who pays to act on it.

The practical response is small enough to start this month. If you run an upazila office, a mill, a cold store, a lending desk or a farm app, write one page before the seedbeds go in. Put on it the forecast you will act on, what you will release when it arrives, and who is allowed to release it without asking. Then send that page to the people upstream and downstream of you. A forecast with a named owner is likely to save more grain than a sharper forecast with none.

Seedbeds for the next Boro crop will go in before the year is out. Sometime in April the bulletin will come again, perhaps earlier in the month than it used to, and it will probably be right. Whether a single dryer or a litre of reserved diesel will be waiting behind it isn't something any model can tell you. Somewhere in Sunamganj a farmer will be told to cut a field that is 80 per cent ripe. He will know what is coming, and so will every office above him. Which of them will be obliged to move?