In my experience the only thing that improves estimates is being radical about breaking tasks up into small, well defined pieces. If my estimate is 8 hours I'm thinking a workday, which might mean 3 days or more. Avoid the 8 hours thermocline :-)
This has a cost though - most of the big tasks can only be broken up semi-honestly (as opposed to a totally made up division into subtasks) after a significant initial part is done, perhaps to be thrown away. This means they cannot really be estimated, go away, it's done when it's done. But if the iterations are reasonably short then at least the management can watch the estimates grow exponentially in real time. Maybe it sounds pessimistic but at least it minimizes the incentives to make fantasy estimates.
This has a cost though - most of the big tasks can only be broken up semi-honestly (as opposed to a totally made up division into subtasks) after a significant initial part is done, perhaps to be thrown away. This means they cannot really be estimated, go away, it's done when it's done. But if the iterations are reasonably short then at least the management can watch the estimates grow exponentially in real time. Maybe it sounds pessimistic but at least it minimizes the incentives to make fantasy estimates.