Poets, priests and politicians
Have words to thank for their positions
Words that scream for your submission
And no one’s jamming their transmission
– The Police. De Do Do Do, De Da Da Da
I am continuing to use R to analyse company announcements. After my last blog post, everyone told me that it was interesting… but that my sample size of 50 companies was far too small. So for my next step, I made my sample size even smaller: to a single company.
There is a method to this. Most of the really big price movements don’t come from 50 page Full Year Results Announcements. Instead they come from “Trading Statements” normally only a couple of sentences long, saying whether the company is going to achieve what investors are expecting – or better or worse.
New information versus prior probability
The ancient Greeks understood that the more surprising an event is, the more information that event carries. If I live in Bergen, in Norway and my local weather forecast tells me it is going to rain at some point during the next week – that’s not surprising. The forecast contains very little information. But a rain forecast in the Atacama desert, where it hasn’t rained for decades is far more informative. You can’t separate new information from your prior probability.
The modern definition of information was formalised by Claude Shannon. The information content of an event is proportional to the log of the inverse probability of it occurring. When you have a lot of uncertainty in your prior probability (whether it’s going to rain next Tuesday at 2pm in Bergen) then a small amount of information (perhaps it rains less frequently in winter than summer) that resolves that uncertainty is highly valuable.
Information =log 1/probability
So I looked at 52 company announcements (18 Full Year, 17 Half Year and 17 Trading Statements) going back to 2000, for a single company. And I manually labelled them depending on whether the share price was 10% or more higher (positive) or if the share price was 10% or more lower (negative) a month later. In all there were 9 negative statements, 20 neutral (ie within a band of less than 10% higher or lower) and 23 were positive.
Meaningless and all that’s true
If I could automate the process, taking 50 companies and analysing 50 announcements from each would give me a sample size of 2,500 which is more big data-ish. I actually have 20,000 announcements, but most are meaningless noise. Companies like to put positive announcements that mean nothing, like small contract wins. I haven’t deleted them yet. I was thinking that there may be signal in this noise, companies that generate meaningless noise probably underperform – if they had a good story to tell, they would tell it.
Solid State Plc IPO’ed in the 1990s at a price of 70p. At the start of 2000 this companies share price was 24p.* By March 2009 the share price had halved to 12p per share. At that point you probably wouldn’t have thought much of this company, waiting for 9 years to lose half your money. If you’d bought in and held from original IPO it would have taken a decade and a half to lose 80% of your money. A bad company or a bad investment?
But 6 years later it reached a peak of 880p. That’s 73 times higher. Good company and a great investment?
Click on the chart to enlarge I’ve put it in log scale because obviously a share price range of 12p to 880p is easier to understand in log scale.
Only cheques I’ve left unsigned
So I’m not just analysing statements to find “good companies”. I also need to identify the correct time to buy. If anyone bought £15,000 worth at 12p per share, they would have been a paper millionaire 6 years later.
But this is not bitcoin – this company was founded in 1971, and has been listed on AIM (the UK regulated stock exchange). It makes stuff: rugged and industrial computers, custom lithium batteries that can last 3 days between charging, secure communication systems and antennas. They sell their product for a profit. I own some (though sadly I didn’t buy at 12p a share).
Aside from this problem – I faced another dilemma. Quants and computer scientists often devise and backtest strategies assuming that the price on the screen is the price that you can deal in size. For liquid currencies like US dollar v Euro or commodities like oil that’s a fair assumption. For small companies, it isn’t. The price on the screen is not deep – people with a computer science background don’t necessarily understand how market making works.
In fact most of the big moves in a share price can happen overnight when the market is CLOSED. Take for example this:
The day before this price sensitive announcement no shares changed hands, though the price quoted was 18.5 pence. Then the Trading Statement came out at 7AM when the market was closed, the market maker saw lots of “Indications of Interest” before the market opened and immediately marked the price up to 28.1 pence, which was also the low of the day. 9,511 shares changed hands and a price at least as high as 28.1 pence. That is a 52% jump while the market is closed.
Training a computer to recognise that this Trading Statement was very price sensitive is one thing, making money from that insight is harder. Interestingly the Trading Statement at the start of May 2009, only sounds cautiously optimistic.
“The strong start to the year reported in our interim results continued through to the end of this fiscal period and we closed FY08/09 with record sales. Although our new order intake slowed in Q4 particularly in components our sound order backlog meant revenues increased sequentially and exceeded expectations given the current economic downturn. Order intake across the Group improved in April and the Board is pleased to report that on 1st May we were awarded a £1.2m order from a leading global defence security and aerospace company for command and control systems for the defence market. £800,000 of this order will be shipped during Q2 with the balance shipping in Q1 of FY10/11. Despite order visibility remaining difficult our strong order book and new product introductions mean the directors remain cautiously optimistic for the year ending 31st March 2010.”
Everyone is supposed to have access to the same information at the same time – but practically market makers need the advantage of being aware of price sensitive information before the market opens, otherwise no one would make any money market making.
High Frequency Trading algorithms work for commodities and currencies, but they don’t work for small companies in equity markets. The bid offer spreads are too wide, and the price sensitive news comes out when the market is closed.
Jamming their transmission
If everyone has access to the same information, and for every buyer there is a seller, how is it possible that some investors do so much better than others? The problem is similar to cryptography, something that Shannon also worked on during the war. The enemy has the same encrypted text, the enemy may even know the encryption system being used. The one thing that the enemy doesn’t have, which the legitimate receiver does… is the key. A secure system works when the enemy has the message and the system, but lacks the key, the systematic procedure to substitute the encrypted text which looks like random noise to meaningful information. I don’t think that it is coincidence that Shannon got interested in the stockmarket and concentrated his holdings into companies with large amounts of upside and long term potential (for instance Hewlett Packard and Teledyne).
Traditional value investing only looks at some simple ratios to determine whether a company is cheap. But someone with imagination and knowledge like Shannon could interpret the same information as everyone else, and conclude that he should put a large proportion of his wealth into buying the companies shares.
Perhaps a machine learning technique could emulate Shannon’s investing style, by identify when well capitalised, profitable companies are still paying dividends and are trading on less than 2x H1 annualised Earnings per Share (H1 diluted EPS 3.9p) versus share price low in March 2009 of 12p. Moreover that when the numbers are OK, but that management are cautiously warning about difficult markets and the valuation is absurdly beaten up – this is the time to really take a large, high conviction position. That is, the price is reflecting a huge amount of uncertainty, but it doesn’t take much new information (a few sentence Trading Statement) to considerably reduce that uncertainty. That might get you a 5 bagger.
Words that scream
In fact it shouldn’t take a computer or a genius like Shannon to recognise that Solid State in 2009 was a buying opportunity. Most people would suggest that the shares at 12p were a screaming buy. Maybe an investor would think the shares would rise to 30p. Maybe even 60p. But perhaps a computer can give you confidence just how big an opportunity Solid State would become: a 73 bagger. I am guessing, but I think the reason Solid State hit such a low valuation was that a large fund manager who owned the shares was seeing redemptions from his fund, and had to sell the shares whatever the price. But if you don’t trust a fund manager and follow your own process, invest your own money you can benefit from this as a private investor.
I see that these guys at Prattle are trying to sell their text analysis software to fund managers. Maybe. I’m not sure that machine learning techniques will solve the problem of fund manager redemptions. Instead I have a better idea. Perhaps technology will make it easier for private investors to improve their process and beat the professionally managed large funds. Small investors with computers have an advantage: they don’t have to worry about redemptions, and they can take large positions in small companies. They can also wait patiently with their cash. I’m working on something so that I (and perhaps others) can process thousands of announcements, waiting for words to thank for our positions.
* Adjusted closing price – adjusted for stock splits, dividends, issuance etc