Fully agree Mark . I'm only using it to gauge the market. I have sent loads of time on 30min to 1min on matchbook look for signals . I will share something later when i get 5min
I’ve been working on a separate project looking at
Matchbook price movement before the off, rather than trying to predict the winner of the race.
The basic idea is to treat the exchange price as a time series. I capture the Matchbook back and lay prices repeatedly from roughly 25 minutes before the race down towards the off, together with things such as spread, market rank and minutes-to-off.
The original question was fairly simple:
can I identify horses which are likely to shorten or drift enough to trade them rather than back them as selections?
I tested both directions.
For a
lay-then-back trade, the idea is to lay a horse first and buy it back later if the price drifts. For a
back-then-lay trade, I back first and try to lay it at a shorter price later. I have been testing different entry times, target movements and market-rank restrictions rather than just looking at the results after the event.
An early replay looked encouraging. Over 47 days, restricting trades to the horses where the model was most confident produced progressively better figures as the probability threshold increased. For example, the liability ROI rose from around
0.39% at a 0.65 model threshold to 1.87% at 0.85. That initially suggested there might be a genuine exploitable relationship between the model probability and subsequent price movement.
The important part, though, was what happened when I tightened the test and made the chronology more realistic.
Once I treated it as a genuine forward-trading problem rather than simply replaying the historical prices, the results became much more modest.
For the
lay-then-back side, most combinations were slightly negative. One small Rank-1 pocket around a 26-minute entry and roughly a 10-tick target was positive, but only marginally so and on too few trades to regard it as a reliable edge.
The
back-then-lay test was much worse. Simply backing a horse around 25 minutes before the off and waiting for a fixed shortening target did not work. That version lost quite heavily.
That has actually been useful, because it has changed the question I am asking.
I no longer think the useful signal is simply:
“This horse will shorten” or “this horse will drift.”
The more interesting information seems to be in the
shape of the move.
For example, some horses do very little from 25 to 15 minutes, then shorten rapidly inside the final 5–10 minutes. Others shorten steadily, look like ideal trades, and then reverse sharply close to the off.
So the next stage is to stop forcing every trade to enter at the same point and instead look at the price path in sections:
25m → 15m early direction
15m → 10m confirmation
10m → 5m acceleration
5m → 1m late move or reversal
I want to measure things such as
persistence, acceleration, spread, reversal and market rank, and then allow the model to decide whether there is actually a trade and when the entry should occur.
The biggest lesson so far is probably that an apparently profitable price-movement model can disappear very quickly once the test is made genuinely chronological. The early results were interesting enough to continue, but the stricter work says there is
not yet a simple “back at 25 minutes” or “lay at 25 minutes” system.
What may still be there is a more subtle edge in recognising
how a price move is developing before the rest of the market has fully reacted. That is what I’m testing next.
You could add one of these details
- Make it more technical
- Add the key results table
Rewrite the forum post in a more technical style, explaining the chronological testing, entry rules, target movements, and why the stricter forward test matters.
I’ve been working on a separate exchange-trading project using
Matchbook pre-race prices, with the aim of predicting short-term price movement rather than the winner of the race.
The data are captured repeatedly before the off, so each runner has a price path through time rather than just one snapshot. The main variables include the Matchbook back and lay prices, spread, market rank, minutes-to-off and the model’s estimated probability of a profitable trade.
The original trading framework was deliberately simple. Entry was targeted at roughly
25 minutes before the off, using a narrow selection window around that point, with the runner generally restricted to the top two in the market and a maximum lay price. From there I tested fixed target movements and a compulsory exit close to the off.
For the lay-to-back side, the structure was:
Enter by laying at around 25 minutes
↓
Wait for the price to drift
↓
Back at the target price
↓
If the target is not reached, close the position near 1 minute to the off
I tested different target distances, effectively asking how often the horse moved far enough in the required direction after entry to make the trade viable after commission and liability.
I also tested the reverse:
Back first
↓
Wait for the horse to shorten
↓
Lay at the target price
↓
Otherwise close near the off
The important part of the work has been the testing method.
An early replay over 47 days looked quite encouraging. As the model probability threshold was increased, the apparent liability ROI improved. At the highest threshold the return was materially better than at the lower thresholds, which suggested that the model’s confidence score was separating stronger and weaker price-movement opportunities.
However, replaying historical data can flatter a trading model if the chronology is not handled very strictly.
The stricter test therefore used a genuine forward-style process. For each trade, the decision had to be made using only the information that would actually have been available at that point in time. The later prices were then used purely to determine whether the target was reached or whether the trade had to be closed out.
That sounds obvious, but it is absolutely critical in this type of work. If later information accidentally leaks into the entry decision, even indirectly, a price-movement strategy can look much better than it really is.
Once I tightened the chronology and re-ran the tests, the picture changed.
The lay-to-back model became much less impressive. Across the broader Rank ≤2 tests, the combinations were generally slightly negative. There was one small positive pocket among Rank-1 runners at around a 26-minute entry and a 10-tick target, but the sample was only 66 trades and the ROI was only about
+0.11%. That is nowhere near enough evidence to treat it as a genuine edge.
The back-to-lay test was weaker again. A simple rule based on backing around 25 minutes and waiting for a fixed shortening target produced a substantial loss over the larger sample.
That result has changed the direction of the research.
The problem with a fixed-entry/fixed-target approach is that it assumes the important information is contained in the price at one arbitrary point in time. The data are suggesting that this is probably too crude.
Some runners do very little between 25 and 15 minutes, then make a strong move late. Others shorten steadily, appear to be ideal back-to-lay candidates, and then reverse sharply in the final few minutes. A fixed 25-minute entry treats both of those price paths in much the same way, even though they are clearly very different market behaviours.
The next stage is therefore to model the
structure of the move, rather than simply the direction.
I am splitting the pre-race period into sections such as:
25m → 15m
15m → 10m
10m → 5m
5m → 1m
and looking at variables such as:
- direction of movement in each interval
- magnitude of the move
- whether the move is accelerating or slowing
- persistence across consecutive intervals
- spread behaviour
- market rank
- reversals
- how late in the market the real move begins
The aim is to move from:
“This horse should drift 10 ticks from 25 minutes out”
to something more like:
“The early move is weak, the 15-to-10-minute move confirms the direction, the 10-to-5-minute move is accelerating and the spread remains tight, so now there may be a trade.”
That is a much harder problem, but it is also a more realistic description of how an exchange market actually develops.
The main conclusion so far is that the early replay was useful, but the stricter chronological test is the one that matters. At present I would not claim to have found a robust fixed-rule Matchbook system. What the testing has shown is that there may be information in the
timing, persistence and acceleration of the price move, but it needs to be modelled dynamically rather than reduced to a single entry time and target.
