HtAG predictions

Hi Rex, see FAQ section on our website for information about accuracy.

 

Thanks. Can't reveal details about source of timely sales data, because it is at the core of company IP. In simple terms, we tap into a number of BigData sources.

Backtesting was done over 10 years worth of sales data. That amounts to around 4,000,000 sales.

 

Hi qak, Is your question for houses or apartments?

 

If your model uses recent sales history, wouldn't it have predicted continued gangbuster gains in many Sydney suburbs a year ago when Sydney, as a whole, was still going gangbusters?

 

So your Algo is extrapolating future price based on what happened before?
So if more recent sale data starts showing decline in price, the prediction will start showing bigger potential declines for future?


How does your model factor in the impact of,

• Current Credit tightening potentially the new normal now,
• 120bn IO2PI cliff for each year uptill 2021,
• Impact of recent credit collation and its effect.
• Upcoming OTP settlements,
• Upcoming surplus supply about to hit market in next two years due to construction boom in good times.
• TotalDebt2income cap (6/7x/?)
• Potential Negative gearing changes? yet uncertain but probable.
• Rising funding costs
• Rising international yields, Systemic risk built up
• Lack of monetary tools at the hand of RBA this time around
• Impact of negative equity due to falling prices and corresponding impact on economy?
• Inability of many investors (even if they wanted to) to be able to buy more at will due to being closer to debt Cap enforced

We have too many headwinds this time around with many coming all at once, would be interesting to see how you guys take this in to account?

 

I guess most recent data given highest weightage

 

Sorry - I'm not trying to be skeptical around your modelling at all - I'd actually be very keen to see things bounce back as per the Hornsby example. I just don't understand how the machine learning adapts to new conditions which have not been modelled in the dataset previously (those conditions being the examples that the SackedWiggle outlines above).

 

So 1 cycle?

 

Houses please

 

Only need one question to either spark or squash further interest....

Are you using Classical or Bayesian methods?

 

Indifference, we use a Long Short-Term Memory (LSTM) recurrent neural network.

 

Data shows that at LGA level, cycles do not move in standard 8-10 year cycles that industry purports. Cycles can last anywhere between 2 to 7 years. All of the areas we report on have finished at least one cycle. Majority are about to complete their 2nd cycle and move into the 3rd.

 

Sam,

The model does not use any of the potential impacts you listed as explicit inputs. Question for you: Do you think that general awareness of these future changes impacts the prices now?

 

That method, often used for non linear time series data, is still marred by the vanishing gradient issue as LSTM only provides some relief due to the sequential data process modelling of the RNN.

Attention-based models are currently preferred as they can look further through the data without needing sequential processing. Ie. Hierarchical neural attention

So why use a LSTM RNN for this data problem when we know that some events are very far apart & infrequent when other techniques are arguably more suited for such data sets?

 

What are you trying to predict... Future or Past?

 

When people can’t get bank loans large enough, it doesn’t matter if they are aware of the inputs.... they are going to be affected...

 

Credit tightening effects price now,
120 bn IO expiry during 2018 effects price now,
Inability to settle OTP during 2018 effects price now.

Coming to future Prediction:
I am very curious to know, what goes in your model to predict future price,
If not Credit availability, forced sales and supply surplus?

 

Hi Mods (@geoffw @Simon Hampel ) @HtAG pinged me thinking I moved the thread. They cannot post as there are restrictions in this forum. Can you have a look?

Also, might be useful to make it clearer that the new thread was not started by the person who's post is first.

 

I'm not a very academic person but to me this all just fluff and speculation, is it just me?? . I'm sure the models mentioned are all correct in theory but the real estate market has multiple factors

To me this is no different to models predicting what the next roulette number is

 

House
Toowong
Qld
4066
Thanks